You can now see how much content, in addition to your prompt, is being sent in your AI requests and easily include or exclude recent messages, helping you control cost and response time.
GPT for Word now warns you when the selected content exceeds the model's maximum context size. This prevents you from accidentally losing important content.
Fixed an issue where copy-pasting content into the chat field didn't work when plain text mode was disabled.
You can now test your Ollama endpoint directly in GPT for Excel and GPT for Word.
When setting up a team, space owners and admins can now choose whether team members are allowed to use their own API keys and, in GPT for Excel and GPT for Word, their own custom endpoints.
Space owners can now transfer ownership directly from the GPT for Work dashboard.
In the GPT for Work dashboard, a new admin role is available to assist space owners with management tasks, including pack purchases.
The progress tracker for bulk AI tools now shows the model used and a runtime counter while a tool is running.
In the Translate bulk AI tool, you can now set a different target language for each row using a separate column.
Result column selection in bulk AI tools is now smarter: The tool automatically suggests the first empty column after your input data.
We’ve restored the default model behavior for the Vision tool in Sheets. It now automatically uses gpt-4o, even if another model is selected.
Connecting to your own OpenAI-compatible endpoints in Excel and Word is now easier and more flexible thanks to a redesigned Custom endpoints section. You can:
Add and manage multiple custom endpoints instead of just one
Check endpoint validity before saving
Provide an optional API key for authenticated endpoints
Add a custom display name to help identify endpoints in the model switcher
Instantly clear all fields and reset your bulk AI tool setup with the New run button.
Inserting content in the Custom prompt bulk AI tool is now easier. You can directly import prompts from cells, and edit them in the GPT for Excel sidebar.
We’ve updated the Generate tool layout to improve clarity. You’ll now find more intuitive field labels and dropdown menus that help you configure each step of the tool more easily.
Fixed an issue where bulk AI tools in GPT for Excel failed to recognize input when only one row contained a value. You can now run executions even with a single populated row.
Today, we're making two important changes to GPT for Work pricing:
Huge price drops for models available without an API key & paid models available with an API key.
Previously free models available with an API key now cost $1 per million tokens.
Models available through custom endpoints (Azure, Ollama, other local servers) remain free for personal use.
For more information, see our blog post.
The progress tracker now shows the model used and a run timer while the bulk AI tools are running.
AI responses that exceed cell character limits in Sheets (50,000) and Excel (32,767) are now automatically shortened and start with a note indicating that the content was truncated, so you always see results instead of an empty cell.
Fixed an issue where using empty global instructions with Anthropic models caused errors in GPT for Excel and GPT for Sheets.
Fixed an issue where the prompt field was disabled if you had an Azure model selected and your GPT for Word balance was empty. The prompt field is now always enabled regardless of your balance since Azure models are currently free.
Fixed an issue where the model switcher did not display Azure models when Azure was the only AI provider you had configured on the API keys & endpoints screen.
You can now use the Anthropic Claude Sonnet 4 model in GPT for Excel and Sheets, both with and without an API key.
Sonnet 4 is a significant upgrade to Sonnet 3.7, delivering superior reasoning while responding more precisely to your instructions.
You can now use the Anthropic Claude Sonnet 4 and Claude Opus 4 models in GPT for Word and Docs, both with and without an API key.
Sonnet 4 is a significant upgrade to Sonnet 3.7, delivering superior reasoning while responding more precisely to your instructions.
Opus 4 is a cutting-edge model with sustained performance on complex, long-running tasks.
We simplified the Custom prompt bulk AI tool by removing two rarely used settings: Custom instructions and Get 2 variants per row.
Fixed an issue where the Reset parameters button in Extract, Summarize, and Reformat bulk AI tools mistakenly triggered the tool instead of resetting the Model settings.
Fixed an issue where the first GPT function executed in a workbook would get stuck on #BUSY!
. This only happened if the add-in wasn't opened beforehand.
You can now use the Anthropic Claude 3.7 Sonnet model without an API key in GPT for Docs.
Claude 3.7 Sonnet is a hybrid-reasoning model that excels both at quick responses and at extended, step-by-step thinking.
You can now invite users with long domain names to your space without hitting validation errors. Previously, invites to domains over 6 characters long would throw an error.
Resolved an issue where GPT_VISION returned cached values despite cache being turned off.
GPT for Docs is moving out of beta to a paid model, using the same billing and pricing as GPT for Sheets. The packs you buy can now be used with both add-ons.
You can now also use GPT for Docs without an API key. Learn more.
You can now choose the image resolution in the Analyze images bulk AI tool, with three options: Low, High, or Auto (default).
Web search with Google Gemini 2.5 Flash is now available without an API key. Web search is priced at $35/1K searches.
Fixed an issue in the GPT for Excel and Word dashboard where space owners could not edit billing info and see and download invoices.
Fixed an issue in GPT for Sheets where adding a valid API key did not display the corresponding models in the model switcher.
We have made GPT for Excel faster for large spreadsheets (10,000+ rows) by reducing the processing time before the first execution by about 50%. The execution time remains the same.
Models added with an OpenAI API key now appear correctly in the model selection menu.
Sonar models in GPT for Excel now correctly limit sources to the domains you select.
The GPT_VISION function now correctly accepts cut-off limit values greater than 16384 with models that support such values.
If you delete an API key from the dashboard, and if your currently selected model in GPT for Sheets is associated with the deleted key, running a bulk AI tool now correctly fails with an error message indicating that the API key is not available.
Fixed an issue where setting an API key failed for all providers and endpoints except for OpenAI.
You can now use the Google Gemini 2.5 Flash model in GPT for Excel, Sheets, and Word, both with and without an API key.
Gemini 2.5 Flash is a cost-effective reasoning model ideal for lightweight tasks and rapid responses.
By default, the model's reasoning capabilities are disabled. You can enable reasoning ("thinking") from the model settings and select between three levels of reasoning effort: Low, Medium, and High.
The Gemini 2.0 Flash model is no longer supported without an API key.
In the dashboard, space owners can now set space-level API keys that are automatically available to all members of the space. Members can then select models available with the keys from the model switcher in GPT for Excel and Word. Members can override space-wide keys by setting their own personal keys.
You can now use AI models available through Azure in GPT for Excel and Word. Define the Azure endpoint URL, model ID, and API key in the API keys & endpoints screen. You can then find and select the model in the model switcher.
Bulk AI tools now show a simplified run status message, without the skipped rows count. The simpler status avoids confusion and helps you focus on the number of rows actually processed.
Fixed an issue where, upon reopening the add-on, GPT for Sheets automatically switched your model to GPT-4o if no API key was set. Your model choice is now preserved across add-on sessions.
Managing API keys is now easier for teams. Space owners can set keys (one per AI provider) that work for everyone in the space. Users can still override the space keys with their own keys if needed.
If your email address changes (for example, due to a domain update), the new email address now appears correctly in all sections of the GPT for Work dashboard.
Custom prompts in GPT for Excel now correctly run even if you accidentally add an extra curly brace ({
) near a variable.
You can now use the following xAI Grok 3 Fast models in GPT for Excel, Sheets, and Word, both with and without an API key:
grok-3-fast (beta)
grok-3-mini-fast (beta)
Grok 3 Fast offers significantly faster response times for latency-sensitive tasks, while Grok 3 Mini Fast is a smaller, cost-effective model optimized for quick and efficient outputs.
Optimized row processing for bulk AI tools, resulting in reduced wait times and fewer output column errors.
Fixed an issue with the Generate bulk tool where the Generate for each row option did not use the Custom content type, when specified, but instead generated content without following your input.
Fixed an issue that caused a blank sidebar to appear when opening GPT for Docs.
You can now use Bulk AI tools on a spreadsheet that was shared with you. Unlike GPT functions, which use the spreadsheet owner's balance and settings, Bulk AI tools use your own balance and settings when you run them as an editor.
You can now use the OpenAI o3 and o4-mini models in GPT for Excel and Sheets, both with and without an API key.
The o3 model is currently the most advanced and powerful reasoning model offered by OpenAI, while the o4-mini model is a smaller, high-performing model optimized for fast and cost-efficient reasoning.
The new Rewrite bulk AI tool in GPT for Sheets allows you to easily rewrite the content of an entire column at once. You can customize instructions, choose content types, and use additional columns for context to get more precise results.
Fixed an issue where GPT formula cache did not work correctly with Anthropic, Gemini, and Perplexity models.
Fixed an issue where the Generate formula assistant displayed the wrong error message when your balance was empty.
The model switcher now displays the Buy a pack button also for Perplexity models when your balance is empty.
The Custom prompt bulk AI tool now supports input columns with trailing spaces in their names. Previously, trailing spaces caused an error and required manual cleanup.
You can now use the OpenAI o4-mini model in GPT for Word, both with and without an API key.
The o4-mini model is a small, high-performing model optimized for fast and cost-efficient reasoning.
You can now use the following OpenAI GPT-4.1 models in GPT for Excel, Sheets, and Word, both with and without an API key:
GPT-4.1
GPT-4.1 mini
GPT-4.1 nano
The GPT-4.1 models improve on the GPT-4o models in speed, accuracy, reasoning capabilities, coding performance, context window size, and cost-effectiveness.
You can now use the following xAI Grok 3 models with an API key in GPT for Excel, Sheets, and Word:
grok-3 (beta)
grok-3-mini (beta)
The grok-2 model is no longer supported.
Clicking the currently selected model in the model switcher now closes the model switcher.
GPT for Sheets now supports up to 64,000 output tokens (63,800 effective) with the Anthropic Claude 3.7 Sonnet model. This allows for much longer, more detailed responses.
You can now use the following Perplexity Sonar models without an API key in GPT for Excel, Sheets, and Word:
Sonar Pro
Sonar Reasoning (only in GPT functions with the model parameter)
Sonar Reasoning Pro (only in GPT functions with the model parameter)
Sonar Pro provides more accurate and detailed results with more comprehensive citations than Sonar, while Sonar Reasoning and Sonar Reasoning Pro provide advanced reasoning capabilities powered by DeepSeek R1.
We increased the price of the Sonar modal when used without an API key:
Tokens: $20.00 per 1 million tokens (previously $2.00)
Searches: $5.00 to $12.00 per 1K searches, depending on the search effort selected (previously $10.00 per 1K searches)
We increased the price of the following OpenAI models:
gpt-4o-mini (without an API key): $10.00 per 1 million tokens (previously $0.90)
o3-mini (without an API key): $10.00 per 1 million tokens (previously $9.40)
The price updates do not affect packs purchased before April 14, 2025.
Fixed an issue where, if you revoked the Perplexity API key that you had set in GPT for Excel, reloading the add-in caused Use your API key models to become unavailable in the model switcher:
If you selected a Sonar model and then reloaded the add-in, all Use your API key models disappeared from the model switcher.
If you selected any model other than a Sonar model and then reloaded the add-in, all Sonar models under Use your API key became unselectable.
Fixed an issue in the Generate bulk AI tool where the search options for Perplexity Sonar models were not displayed.
Fixed an issue where, if your balance was empty when you opened the add-on, clicking the back button quickly on the Billing page navigated you back to the same page.
The examples in the GPT functions list now use the correct argument separator (colon or semicolon) for your Google Sheets locale, so the examples no longer trigger errors when copied and inserted into the spreadsheet.
Fixed an issue in the Generate formula assistant where syntax highlighting for the generated formula caused GPT for Sheets to freeze.
You can now use the following Mistral models with an API key in GPT for Word:
Mistral Large
Mistral Small
Ministral 8B
Ministral 3B
Mistral models prioritize cost-efficiency and open-source accessibility, and provide a balanced solution for a wide range of applications.
Fixed an issue on macOS where, if you used the Japanese Input Method Editor (IME), pressing ENTER immediately submitted the prompt instead of opening the prediction candidate window. Pressing ENTER now allows you to select a typeahead candidate, while you submit the prompt by clicking Send.
You can now use the following Perplexity Sonar models with an API key in GPT for Excel and Word:
Sonar Pro
Sonar Reasoning
Sonar Reasoning Pro
Sonar Pro provides more accurate and detailed results with more comprehensive citations than Sonar, while Sonar Reasoning and Sonar Reasoning Pro provide advanced reasoning capabilities powered by DeepSeek R1.
All Sonar models used with an API key now cost $5 per 1 million tokens.
You can now set the Search effort for Sonar models when using an API key. The setting allows you to balance cost against search result quality.
You can now use the following Mistral models with an API key in GPT for Excel:
Mistral Large
Mistral Small
Ministral 8B
Ministral 3B
Mistral models prioritize cost-efficiency and open-source accessibility, and provide a balanced solution for a wide range of applications.
We increased the price of the following OpenAI models:
gpt-4o-mini (without an API key): $10.00 per 1 million tokens (previously $0.90)
o3-mini (without an API key): $10.00 per 1 million tokens (previously $9.40)
The price updates do not affect packs purchased before April 9, 2025.
Fixed an issue where using a selection or document as context resulted in an error when GPT for Word tried, under the hood, to expand the selection to cover leading or trailing parts, such as the first word of a paragraph, that you had likely omitted by mistake.
This issue only occurred on older versions of desktop Word.
You can now use the Google Gemini 2.0 Flash model without an API key in GPT for Excel, Sheets, and Word. The model offers fast results, a large context window, advanced multimodal capabilities, and cost-effective performance.
The Google Gemini 2.5 Pro Preview model replaced the Gemini 2.5 Pro Experimental model.
Fixed an issue in the Generate formula assistant where syntax highlighting for the generated formula caused GPT for Excel to freeze.
You can now use fine-tuned OpenAI GPT-4o and GPT-4o mini models with an API key in GPT for Word. Fine-tuning allows you to train base models with your own data, tailoring them for your specific needs and applications.
You can now use fine-tuned OpenAI GPT-4o and GPT-4o mini models with an API key in GPT for Excel. Fine-tuning allows you to train base models with your own data, tailoring them for your specific needs and applications.
Bulk AI tool runs now clearly show skipped rows in GPT for Sheets. Rows are skipped when:
Content is found in the result column
All input columns are empty
This change helps bulk AI tools run faster.
You can now use the following Google Gemini 2 models with an API key in GPT for Word:
Gemini 2.0 Flash
Gemini 2.0 Flash Lite
Gemini 2.5 Pro Experimental
Gemini 2 models offer fast results, low costs, and advanced reasoning.
When you select multiple paragraphs in GPT for Word, the full beginning and end of the selection are now automatically included in the context to avoid accidental cutoffs.
You can now use the following Perplexity Sonar models with an API key in GPT for Sheets:
Sonar Pro
Sonar Reasoning
Sonar Reasoning Pro
Sonar Pro provides more accurate and detailed results with more comprehensive citations than Sonar, while Sonar Reasoning and Sonar Reasoning Pro provide advanced reasoning capabilities powered by DeepSeek R1.
All Sonar models used with an API key now cost $5 per 1 million tokens.
You can now set the Search effort for Sonar models when using an API key. The setting allows you to balance cost against search result quality.
The new Generate bulk AI tool in GPT for Sheets makes it easier to generate text based on your spreadsheet data, with options to choose content types and use multiple columns for context.
You can now use the following Google Gemini 2 models with an API key in GPT for Excel:
Gemini 2.0 Flash
Gemini 2.0 Flash Lite
Gemini 2.5 Pro Experimental
Gemini 2 models offer fast results, low costs, and advanced reasoning.
You can now use cell formatted as Date or Time as input in GPT functions like: =GPT(“Convert this date to ISO”, B1)
You can now use the following Google Gemini 2 models with an API key in GPT for Sheets:
Gemini 2.0 Flash
Gemini 2.0 Flash Lite
Gemini 2.5 Pro Experimental
Gemini 2 models offer fast results, low costs, and advanced reasoning.
The Google Gemini 1.5 models are no longer supported.
If you're signed out of the GPT for Excel or Word add-in, the sign in screen now appears as expected.
We've fixed an issue where a blank OpenRouter model appeared in the model switcher.
We fixed an issue where the API key fields in GPT for Excel were empty even though a key was already saved.
In the settings screen, the Reset to default button in the Custom instructions section now appears correctly, even when you select the Empty preset and delete all text.
The Custom prompt bulk tool no longer crashes when your spreadsheet column headers contain numbers like 2,1 or 12,4. You can now safely use headers with comma-separated numbers in both GPT for Sheets and GPT for Excel.
You can now set the following search options in bulk AI tools when using the Perplexity Sonar model:
Domains: Specify up to three domains to target. The model will attempt to generate responses based only on search results from these domains.
Recency limit: Select a time range for the search results. The model will generate responses using only results from this period.
Global instructions now correctly process empty inputs instead of falling back to the default value.
GPT for Word now supports OpenRouter, giving you access to over 300 AI models through a single API. Use an API key to get started and access nearly any open-source model and more.
Web search in GPT for Excel just got better: gpt-4o-search-preview and gpt-4o-mini-search-preview models are now available alongside Perplexity's sonar, giving you more choice for real-time information. You can adjust the models' context size and location parameters directly in the bulk AI tools.
GPT for Excel now supports OpenRouter, giving you access to over 300 AI models through a single API. Use an API key to get started and access nearly any open-source model and more.
Bulk AI tool runs now clearly show skipped rows in GPT for Excel. Rows are skipped when:
Content is found in the result column
All input columns are empty
This change helps bulk AI tools run faster.
Fixed an issue where the sidebar footer did not show your real-time balance.
You can now set the following search options in bulk tools when using the Perplexity Sonar model:
Domains: Specify up to three domains to target. The model will attempt to generate responses based only on search results from these domains.
Recency limit: Select a time range for the search results. The model will generate responses using only results from this period.
Copying a portion of text in the Custom prompt bulk tool field now correctly copies only the selected text instead of the entire field.
Chat history is now automatically managed to prevent slowdowns. When history exceeds 200 messages or 25,000 characters, old messages are deleted while keeping at least 20 messages or 5,000 characters to maintain context.
HTML source code no longer appears in chat messages when restarting the add-in with Plain text mode enabled.
You can now define dedicated instructions for the Custom prompt bulk tool. These instructions are applied on top of the global custom instructions defined in the spreadsheet settings.
Global instructions for bulk tools now correctly restore when reopening saved instructions from history. Previously, the instructions section remained empty.
In GPT functions, models that use an API key are now automatically prioritized when specifying a model with the model parameter.
Custom prompt bulk tool instructions now restore correctly from history. Previously, they displayed the latest updated instructions.
All members of a team now have access to the team balance and space details in the add-ons. Previously, only the owner could access them.
You can now also use Perplexity AI's sonar model in GPT for Excel and GPT for Word with an API key.
Bulk tools now correctly send the displayed text value to AI models, preventing confusion when working with dates, currency, and other formatted data.
The API keys page no longer includes the OpenAI organization ID field, making setup and key management more straightforward.
The Generate and Rewrite bulk tools now handle missing fields correctly when revisiting bulk tool history, preventing crashes and displaying validation errors instead.
GPT for Excel now correctly recognizes and counts active GPT formulas, even when nested inside other formulas.
GPT-4.5-preview is now supported in GPT for Sheets, and in GPT for Excel Word! If you have an OpenAI API key, you can now access OpenAI's latest model, which performs well on tasks that require creative, open-ended thinking—such as writing, learning, and brainstorming new ideas.
Fixed an issue in the Custom prompt tool where variables were not replaced when the corresponding column contained empty cells.
Long instructions in the Classify/Categorize, Extract, and Reformat bulk tools no longer overlap with Spreadsheet settings.
Fixed an issue that caused the Word add-in to crash when sending multiple requests quickly to an Ollama model. This was due to an unexpected change in how events are sent in the latest Ollama release.
You can now use xAI's Grok models in GPT for Word with an API key. The Grok models excel at reasoning and coding tasks. GPT for Word currently supports the grok-2 model.
Long chat histories no longer cause the add-in to slow down or freeze when sending new messages.
Fixed an issue on older Microsoft Word for Mac versions where prompting the AI resulted in the error: An unexpected error occurred. Please try again.
You can now use xAI Grok models in GPT for Excel with an API key. The Grok models excel at reasoning and coding tasks. GPT for Excel currently supports the grok-2 model.
You can now use xAI Grok models in GPT for Sheets with an API key. The Grok models excel at reasoning and coding tasks. GPT for Sheets currently supports the grok-2 model.
You can now use the Anthropic Claude 3.7 Sonnet model in GPT for Sheets with an API key. Claude 3.7 Sonnet is a hybrid-reasoning model that excels both at quick responses and at extended, step-by-step thinking. The model also provides enhanced coding capabilities.
You can now use the Perplexity Sonar model in GPT for Sheets with an API key. For more information about the model, see the February 24 note for GPT for Sheets.
Resolved an issue in bulk tool history where opening a previously saved run did not restore the correct model. The model selection only corrected itself after opening another saved run.
You can now use the OpenAI o3-mini model in GPT for Sheets, both with and without an API key.
o3-mini is a small, cost-efficient reasoning model that delivers advanced reasoning capabilities that are especially well-suited for coding tasks.
o3-mini supports three reasoning effort options: Low, Medium, and High. You can select the option from the spreadsheet settings.
The new Generate bulk tool in GPT for Excel makes it easier to generate text based on your spreadsheet data, with options to choose content types and use multiple columns for context.
Resolved an issue where GPT for Word returned an incorrect error message when the LLM server for a custom endpoint was unavailable.
GPT for Word now supports any locally running model that is compatible with the OpenAI API.
If you have a compatible model running on a local machine and you want to use the model in GPT for Word, set the model's endpoint URL in the API keys & endpoints screen.
The model switcher button in GPT for Excel and Word now shows model names correctly.
You can now see detailed model token costs by hovering over the $$$ icons in the model switcher for GPT for Excel and Word.
Resolved an issue where a blank screen appeared when opening the GPT for Excel Word add-in. The add-in now loads faster with an optimized loading screen.
GPT for Word now supports the following new model settings:
Creativity: Define how standard (or not) your result is expected to be.
Top P: Adjust the creativity level. Lower values result in more focused output, while higher values allow for more creative responses.
Frequency and Presence penalties: Reduce the repetition of tokens in responses with these OpenAI-specific settings.
These settings are not supported for the o1-mini model.
The Perplexity Sonar model sonar replaces the sonar-large and sonar-small models in GPT for Excel.
The Search the web bulk tool and GPT_WEB function now use Perplexity AI's sonar model in GPT for Sheets, replacing the llama-3.1-sonar-small-128k-online model.
Resolved an issue in the model switcher where the Set API keys & endpoints button was missing when the Ollama endpoint was not set.
Resolved an issue where, if the model switcher was open, selecting Usage tracking or Manage API key from the sidebar menu did not close the model switcher, hiding the selected page.
The model switcher in GPT for Excel and Word now includes a search bar. If you have a long list of models, use the search bar to filter the list by AI provider name and model name.
GPT for Excel now supports any locally running model that is compatible with the OpenAI API.
If you have a compatible model running on a local machine and you want to use the model in GPT for Excel, set the model's endpoint URL in the API key & endpoints screen.
You can now use the OpenAI o1 and o1-mini models in GPT for Sheets.
The o1 models deliver advanced reasoning capabilities and are well-suited for coding tasks, mathematics, and scientific research. The o1 model provides high-performance reasoning across a wide range of tasks but at high cost, while the o1-mini offers a cost-efficient alternative with similar reasoning power but at much lower cost.
You can now select the Perplexity Sonar models sonar-large and sonar-small in GPT for Excel. This means you can now search the web from any bulk tool or function! You do not need an API key to use either model.
Since web search is now available through the Perplexity Sonar models, the Search the web bulk tool is no longer supported. The GPT_WEB function remains available for now.
The Prompt images (Vision) bulk tool and the GPT_VISION function always use gpt-4o. You cannot use any other models with them, not even the Perplexity Sonar models.
Added the following buttons for interacting with previous user prompts in the chat history:
Retry: Rerun the prompt.
Copy: Copy the prompt to the clipboard.
Delete: Delete the prompt and its response from the chat history.
The buttons are displayed when you hover over a prompt.
If a response consists of a single table, the chat now shows the Insert, Replace, and Copy buttons only once, instead of showing them both for the table and for the response as a whole.
Resolved an issue where an empty response incorrectly threw an error.
Resolved an issue where the cut-off limit in Safe mode, when set to Auto-replace formulas, was not working correctly and causing responses to truncate.
Resolved an issue where the Auto/Fixed toggle for Cut-off limit was not displayed when Safe mode was enabled and an assistant model based on gpt-4 was selected.
You can now define dedicated instructions for the Custom prompt bulk tool. These instructions are applied on top of the global custom instructions defined in the spreadsheet settings.
Resolved an issue where the chat sometimes got stuck when using the o1-mini model.
Switching accounts after running bulk tools with a $0 balance no longer causes the Buy a pack to continue modal to remain open.
Resolved an issue in GPT for Word where text was inserted in the wrong place when working with merged tables cells.
Resolved an issue that caused incorrect cut-off limits when using OpenAI Assistants in GPT for Sheets. Now, the cut-off limit setting is model-specific.
Fixed an issue where clicking Clear chat would mistakenly erase your prompt in the input box. Your message is now preserved.
Resolved a bug where using an OpenAI Assistant model ID as a custom function argument caused errors or returned responses from the wrong model.
Resolved an issue where bulk tools used the cut-off limit defined for GPT functions, resulting in truncated responses. The default behavior, where bulk tools do not have a cut-off limit, is now restored.
In the model switcher, the model tooltips now display a list of capabilities with ratings. Previously, the tooltips displayed a summary of pros and cons.
You can now manually stop in-progress Insert and Replace operations on formatted text. If a large insertion or replacement is taking too long or making the Word UI sluggish, click Stop to cancel the operation. Any changes already made by the operation will remain.
Improved performance when replacing text that contains list formatting. Replacing list items now works up to 10 times faster than before.
You can now use the OpenAI o1 and o1-mini models in GPT for Excel.
The o1 models deliver advanced reasoning capabilities and are well-suited for coding tasks, mathematics, and scientific research. The o1 model provides high-performance reasoning across a wide range of tasks but at high cost, while the o1-mini offers a cost-efficient alternative with similar reasoning power but at much lower cost.
You can now use the OpenAI o1-mini model in GPT for Word.
The o1-mini model delivers advanced yet cost-effective reasoning capabilities and is well-suited for coding tasks, mathematics, and scientific research.
Resolved an issue on the API key error page where clicking the I've fixed the problem, retry button returned you to the home page even when the problem was not yet fixed. Clicking the button no longer navigates away from the error page if the problem persists.
We fixed an issue where bulk tools executed an empty row when Start from row or Header row was set beyond the last row containing data. Bulk tools now skip all empty rows.
Running bulk tools immediately after modifying settings no longer causes outdated configurations to be applied.
Managing your team just got easier!
You can use the new GPT for Work dashboard to add and remove members to and from your team and buy packs for your team. To access the dashboard, open the add-on main menu and click Admin, or go directly to https://dashboard.gptforwork.com.
Following the popularity of the speed-run videos, we've added a bulk tools walkthrough to the Help tab. Get started with bulk tools in minutes!
Bulk tools no longer process hidden rows when generating results. Bulk tools now only process visible rows.
Resolved an issue where the Classify / Categorize bulk tool incorrectly returned #N/A as the result, even though the AI model provided a valid response.
Resolved an issue where bulk tool runs sometimes got stuck with very long prompts (roughly 100,000 characters or more).
In the Custom prompt and Translate bulk tools, enabling the Get 2 variants per row option now correctly generates two variants instead of one when using Anthropic models, Gemini models, and OpenAI Assistants.
Resolved an issue where the formula assistant sometimes wrapped generated formulas in Markdown.
Upgraded the Anthropic Claude Haiku model from version 3 to version 3.5.
Resolved an issue where the Prompt images (Vision) and Search the web bulk tools and the GPT_VISION and GPT_WEB functions did not consume balance when a free model was selected. These features use specific paid models, ignoring what is selected in the model switcher, so they always consume some balance.
Resolved an issue where the Search the web bulk tool sometimes failed to return sources when running the tool for all rows at once (Run all rows). This bug only affected a small minority of users.
GPT for Sheets now supports additional Mistral models:
mistral-small: Ideal for quick and efficient processing
ministral-8b and ministral-3b: Designed for more complex tasks
You need an API key to use these models.
We've fixed an error that occurred when specifying a model as a parameter in a GPT formula while an OpenAI assistant was selected in the model switcher.
The bulk tools Classify / Categorize, Prompt images (Vision), Reformat, and Search the web now generate consistently formatted responses with greater reliability.
The Start from row tooltip in bulk tools now dynamically updates based on the value of Header row.
You can now cancel bulk tool runs of fewer than 20 rows.
To avoid "Insufficient permissions" errors, GPT for Sheets now checks your permissions and requests sign-in when necessary.
You can now log out of your Microsoft account directly from the GPT for Excel or GPT for Word sidebar. Open the sidebar menu and click Log out.
We've resolved an issue where OpenAI keys with no quota were treated as valid, leading to execution errors. GPT for Excel now displays an Action required button, helping you resolve the issue quickly.
We've fixed an issue where, if your Anthropic API key was revoked, the model switcher still showed Claude models under Use your API key.
Upgraded the Anthropic Claude Haiku model from version 3 to version 3.5.
GPT formulas with the model parameter now prioritize models requiring an API key whenever one is available, even if a “No API key needed” model is selected in the model switcher.
We've fixed an issue that prevented the Extract bulk tool from working with the GPT-4 model.
Bulk tools now ignore hidden rows during execution.
The Search the web bulk tool now correctly excludes sources from its output when Show sources is disabled. Similarly, the GPT_WEB function now excludes sources when the show_sources parameter is undefined or set to false.
Bulk tools now execute successfully even when input cells contain a dollar sign ($) followed by a special character, for example $' or $!.
If you're using an Anthropic Claude model, the AI now correctly identifies itself as Claude from Anthropic rather than as ChatGPT from OpenAI.
We've fixed an issue in responses where the Replace button was shown for tables when Use selection or document for context was enabled but no content was selected for replacement in the document.
The new Reformat bulk tool in GPT for Excel and GPT for Sheets allows you to easily reformat the content of a column. You can, for example, adjust dates to use a specific format, turn text into bullet lists, change text case, and rewrite addresses in a specific format.
This tool focuses on cell content and does not allow you to change cell formatting, such as font or border settings.
To access the tool, select Bulk tools in the sidebar and click Reformat.
The Search the web bulk tool now correctly excludes sources from its output when Show sources is disabled. Similarly, the GPT_WEB function now excludes sources when the show_sources parameter is undefined or set to false.
The Custom prompt bulk tool now correctly inserts column variables at cursor position after line breaks.
The UI of the desktop add-in now remains responsive even with large workbooks containing millions of cells of data.
Bulk tools no longer truncate responses after 500 tokens when used with Anthropic models. Models from other providers were not affected by this bug.
We've fixed an issue where plain text mode was enabled by default. To use plain text mode now, you must again explicitly enable it.
Bulk tools now ignore blank rows that contain one or more spaces. This avoids unnecessary processing and token consumption.
The UI now shows a preview of custom instructions and spreadsheet settings when the corresponding UI elements are collapsed, making the settings more readily visible.
GPT functions now correctly remove backticks that AI models use to wrap responses (response
).
In the Custom prompt bulk tool, you can now easily insert column variables into the prompt by using the column label and column letter drop-down menus. Adjustments to the column name row are reflected instantly in the selection.
Selecting large sections of text, especially formatted text, no longer causes Word to slow down or freeze, even in large documents.
The "OpenAI API isn't working with your key" error message now clears automatically upon entering a valid API key, without the need to change the model.
We've fixed an issue in the bulk tools where setting Start from row to greater than “5000” caused results to be placed in unintended rows.
The new Excel formula assistant can generate formulas from descriptions or explain existing formulas.
Accessible in both the GPT functions and Bulk tools tabs, and billed the same as GPT functions.
We've addressed several issues in the Extract bulk tool:
Extraction now works correctly when rerunning the tool, even if the first result column was left empty during the initial run.
A trailing comma at the end of the entity name in the Extract field no longer impacts extraction results.
You can now easily access and re-use past Bulk tool configurations with our new History feature. Apply past settings in a single click, even across different workbooks.
Set the model of your choice for each GPT function with the new model parameter. This parameter is available for all functions except GPT_WEB, GPT_VISION, and GPT_MATCH.
We've fixed an issue in the Custom prompt bulk tool where the Run 3 rows option required at least four empty cells in the output column to work correctly.
You can now run the Custom Prompt bulk tool on the online version of GPT for Excel with four or more variables in your prompt template.
A new GPT for Sheets speedrun video is now available in the Get started section of the sidebar, providing a quick overview of all GPT functions.
You can now use specific snapshot versions of models in GPT function parameters when working with an API key. Previously, specifying snapshots defaulted to the generic model. For example =GPT("Write a tagline", A2, , "gpt-4o-2024-08-06") no longer defaults to gpt-4o.
In the Custom prompt bulk tool, you can now easily insert column variables into the prompt by using the column label and column letter drop-down menus. Adjustments to the column name row are reflected instantly in the selection.
The column selection in GPT for Excel bulk tools now remains active even after adding or removing rows, without reloading the tool.
You can now easily access and re-use past Bulk tool configurations with our new History feature. Apply past settings in a single click, even across different spreadsheets.
Links with empty href
attributes now appear as plain text. Previously, clicking these links opened a blank or endlessly loading screen.
The Extract bulk tool now correctly outputs values when using a comma-separated list of entities in the Extract field.
We've fixed an issue where retrying errors in bulk tools only cleared cells.
The Keep formatting option is now renamed to Plain text mode, and is disabled by default.
After selecting text in a document and switching to Plain text mode, the selection is now refreshed according to the new mode. This fixes an issue where responses could incorrectly display HTML instead of plain text.
We've resolved an issue where the Extract bulk tool incorrectly formatted results as JSON when an OpenAI Assistant was selected in the model switcher.
Bulk tools now let you set the Header row, ensuring your headers are used in dropdowns for selection.
A system error could cause the sidebar to stop responding after running a bulk tool.
We resolved an issue where the sidebar could display as blank, either when opened manually or when it opened automatically during GPT function execution.
We've fixed an issue in GPT for Word where using Anthropic models with long chat histories could cause an error. Previously, you had to clear the chat history to continue.
Small runs in GPT for Sheets bulk tools temporarily did not return any results. This issue was quickly resolved.
GPT for Word now supports headings inside lists. You can generate and insert these advanced formatting options directly into your Word document via the add-in.
The Custom prompt bulk tool in GPT for Sheets now correctly recognizes column variables, even when nested inside other braces, like {{test{{A}}}}.
We've resolved multiple issues affecting document selection, including unexpected errors and performance slowdowns.
The issue that prevented you from editing long inputs when the output column was in error has been fixed.
You can now open GPT for Sheets even when you have reached Google rate limits. This ensures uninterrupted access to features and settings from the sidebar.
Bulk tools no longer allow column selections beyond the available sheet range, preventing potential selection errors.
We've fixed an issue where large sheets caused the sidebar to crash due to automatic formula counter refreshes. Now, the formula counter only refreshes when manually triggered.
We've addressed an issue in GPT for Word where Anthropic models failed to respond to new messages after interrupting a previous response.
We've improved list formatting in GPT for Word by restricting deep list nesting to 9 levels, matching Word's limits to avoid formatting issues.
In this release, we've improved text handling and formatting in the online version of GPT for Word. Key updates include:
Compatibility with non-English interfaces
Proper handling of text replacement inside paragraphs
Proper text formatting when inserted at pre-formatted locations
In the Custom prompt bulk tool, you can now reference a range from a different sheet. For example: Reference a range from another sheet: {{Sheet2!A1:B5}}.
We have just launched the Prompt images (Vision) tool in GPT for Sheets.
This tool allows you to analyze URLs in a column using AI vision capabilities. To access it, simply click on Prompt images (Vision) from the Bulk tools tab in the GPT for Sheets sidebar.
This tool uses OpenAI's model, GPT-4o, for image analysis.
We've fixed an issue in the bulk tools where the Put results in column dropdown failed to load in large sheets with around 1.5 million cells (approximately 50K rows up to column Z).
GPT for Excel and Word now support locally run open-source models through Ollama integration.
Enter your Ollama endpoint URL in the API key management screen to access local models such as Meta Llama, Google Gemma, Alibaba Qwen, or Mistral Nemo.
The output token limits for Anthropic models have been adjusted to match their actual capacities.
In GPT for Excel, Claude-3-Haiku and Claude-3-Opus now support up to 4096 output tokens, and Claude-3.5-Sonnet supports 8192 output tokens. Previously, the output limit was set to 500 tokens.
The Extract tool in GPT for Sheets now supports multiple extracts in a single execution, allowing you to populate multiple columns per row seamlessly.
Inserting and managing categories in the Classify bulk tool and glossary entries in the Translate bulk tool is now easier. You can directly import lists of categories or glossaries from your spreadsheet, edit them in the GPT for Sheets sidebar, and copy them back into the selected cell with just one click.
Classify bulk tool with the new category management feature:
Translate bulk tool with the new glossary management feature:
Inserting and managing glossary entries in the Translate bulk tool is now easier. You can directly import the content of an existing glossary from your workbook, edit it in the GPT for Excel sidebar, and copy it back into the selected cell with just one click.
We've fixed a bug within Bulk tools where the Retry errors button would run just a couple of rows instead of running all rows previously on error.
We've fixed an issue within Bulk tools where errors in the first column weren't replaced after running out of balance during the execution of a bulk tool with 2 variants per row. The tool now correctly updates all columns with results.
Inserting and managing categories in the Classify bulk tool is now easier. You can directly import an existing list of categories from your workbook, edit it in the GPT for Excel sidebar, and copy it back into the selected cell with just one click.
We've resolved a problem that caused bulk tool executions to load indefinitely when using an API key with a model that is no longer supported.
The Custom prompt and Translate bulk tools now allow 2 variants per row in a single execution, making it easier and cheaper to find the perfect output. This feature is available for all models except OpenAI Assistants and fine-tuned models.
Anthropic's claude-3.5-sonnet model is now available without an API key at $20 per 1M tokens with an input size limit of 200K tokens.
This model is also available with an API key, along with claude-3-haiku and claude-3-opus.
For new users, the Keep formatting feature in GPT for Word is automatically activated, making it easier to maintain their document's original format.
We've resolved multiple issues with AI-generated tables in GPT for Word. These fixes eliminate errors when inserting tables with merged cells.
We've corrected an issue where Safe mode was being automatically enabled when granting permissions to Bulk tools. Now, Safe mode remains off until you choose to enable it.
We've resolved an issue causing the Column is empty error when using the Run 3 rows option in Bulk tools. This occurred when empty cells preceded the values in the column.
The maximum requests per minute (RPM) for the GPT-4o and GPT-4o-mini models has been increased to 1,000, and for all other models to 400, up from the previous limit of 300 across all models. The tokens per minute (TPM) limit is now set to 30 million for the GPT-4o and GPT-4o-mini models, and 2 million for other models.
These new limits apply if you are using the models without an API key. If you are using an API key, your rate limits are still determined by your model provider's organization settings.
When attempting to insert or replace text in a restricted section of a document, you will now get a clear notification instead of an unexpected error.
We've significantly improved the insertion speed when using the Keep formatting option in the online version of Word. Previously, inserting simple content took about 2 seconds. Now, it's almost instantaneous.
You can now conveniently check the current settings for custom instructions, temperature, and top_p values for the selected OpenAI Assistant in the GPT for Sheets sidebar.
We've increased the rate limit for bulk AI tools, with the maximum requests per minute (RPM) now set at 360, up from the previous limit of 180. The tokens per minute (TPM) limit is now set to 30 million for the GPT-4o and GPT-4o-mini models, and 2 million for other models. This update applies only if you do not use an API key. If you are using an API key, your rate limits are still determined by your model provider's organization settings.
We've resolved an issue where unnamed OpenAI Assistants were displayed as "undefined" in GPT for Sheets sidebar settings. Now, the Assistant ID appears correctly, matching the model switcher behavior.
We've fixed an issue that caused an error when using the Regenerate selection feature with multiple rows or columns selected.
The Web browsing bulk tool in GPT for Excel and the GPT_WEB function in GPT for Excel and Sheets now use the Perplexity AI's llama-3.1-sonar-small-128k-online model. Previously, the llama-3-sonar-small-32k-online model was used.
You can now use your OpenAI Assistants directly in GPT for Excel! Set up your OpenAI API key to make your Assistants available in the model switcher.
In the Search the web bulk tool, spaces in the Sources column are now correctly replaced with the source URL instead of being left empty. We've also fixed a bug that previously blocked variant generation in columns containing numbers or booleans.
The issue causing the GPT for Excel add-in to refresh the sign-in with Microsoft page every 2-3 minutes after a manual sign-in has been fixed.
To prevent errors, any model configuration that is no longer supported now defaults to gpt-4o. This improvement applies to models without API keys.
You can now add specific instructions to the Classify and Extract tools, enabling more customized and accurate outputs.
We've enhanced the AI's ability to handle document formatting in GPT for Word. Now, when you select formatted text (e.g., paragraphs, lists, tables), the formatting is preserved in the AI's response. You can insert or copy the formatted text directly into your document.
Anthropic's claude-3.5-sonnet model is now available without an API key at $20 per 1M tokens with an input size limit of 200K tokens.
We've enhanced the Variants per row feature to generate responses that are even more distinct, further increasing variability and usefulness.
We've increased the TPM (tokens per minute) limitation when using providers without an API key (Gemini, Mistral, OpenAI, and Anthropic). Additionally, we've improved rate limit implementations when using Anthropic with an API key.
This is a first step towards further improvements to be announced soon.
Assistants pricing is now based on the underlying model of the assistant, instead of a fixed $5 per million tokens. For example:
Assistants using gpt-4o: $5 per million tokens
Assistants using gpt-4o-mini: $0.3 per million tokens
We have fixed an issue that prevented file search and code interpreter from being used simultaneously. Previously, the code interpreter was disabled whenever file search was enabled, preventing the Assistant from performing calculations after searching a file.
You can now start bulk tool executions from a specific row in GPT for Excel, allowing you to test results on targeted content. For example, you can run the next 5 empty rows starting from row 4.
We've fixed an issue where incorrect success and error messages appeared when your balance was $0.00 or less than $0.05.
You can now use the gemini-1.5-flash model without an API key, which costs $2.50 per 1M tokens.
You can now start bulk tool executions from a specific row in GPT for Sheets, allowing you to test results on targeted content. For example, you can run the next 5 empty rows starting from row 4.
We've resolved an issue where "Replace formulas" caused GPT functions with a prefix to lose it and generate an unintended server request instead of being ignored.
Embedding costs have been reduced:
Without an API key: text-embedding-ada-002 (GPT_MATCH) now costs $0.25 per 1M tokens, previously $2.50.
With an API key: text-embedding-ada-002 (GPT_MATCH) now costs $0.1 per 1M tokens, previously $1.
We've fixed a bug where the "Buy a pack to continue" message was missing when you ran out of credits and tried to run a bulk tool.
The Extract tool in GPT for Excel now supports multiple extracts in a single execution, allowing you to populate multiple columns per row seamlessly.
We've fixed a bug where the modal switcher remained open during bulk tool execution, blocking the view of the execution status.
The Custom prompt bulk tool now features live prompt previews directly under the Prompt field, allowing real-time updates with the data in your sheet.
We are excited to announce that OpenAI's latest model gpt-4o-mini is now supported by all our extensions for Microsoft Excel, Word, Google Sheets, and Docs. It is is smarter than gpt-3.5-turbo and is the most cost-efficient small model:
With an API key, it costs $0.90 per 1M tokens
Without an API key, it costs $0.30 per 1M tokens
Specific instructions are now available for the Classify and Extract bulk tools for more tailored output.
We've resolved an issue where functions with empty cells as parameters incorrectly threw an error. Now, completely empty parameters will prompt an error to set the required parameter, while ranges with some empty cells will ignore the empty cells.
The Custom prompt bulk tool now loads large workbooks more efficiently, enhancing dynamic preview performance.
The Custom prompt and Translate bulk tools now allow 2 variants per row in a single execution.
When generating multiple variants, GPT for Excel enforces a temperature of 0.8 to ensure a variety of high-quality results.
The Search the web bulk tool no longer shows an empty error when you select an OpenAI Assistant in the model switcher.
We have just launched the Search the web tool in GPT for Sheets.
This tool allows you to search the web to get up-to-date information. You can ask questions about events, facts, people, URLs… You can also display the web search sources in a column to understand where the information comes from and assess its reliability.
To access it, simply click on Search the web from the Bulk tools area in the GPT for Sheets sidebar.
The price is $15 or €15 per 1000 executions.
When an OpenAI Assistant is selected in GPT for Sheets, the settings for temperature, top_p, and custom instructions are taken from the OpenAI Assistant configuration, and can no longer be overwritten in the GPT for Sheets sidebar.
We've resolved an issue where the bulk tools failed when used with sheets having personalized names and more than 20 rows.
We have just launched the Prompt images (Vision) tool in GPT for Excel.
This tool allows you to analyze URLs in a column using AI vision capabilities. To access it, simply click on Prompt images (vision) from the Bulk tools area in the GPT for Excel sidebar.
This tool uses OpenAI's latest model, GPT-4o, for image analysis.
We are excited to announce the release of the Classify bulk tool in GPT for Sheets. This new tool allows you to classify the content of a column into predefined categories. For example, classify your user feedback into relevant categories for your Support team. Learn more.
You can now start bulk tool executions from a specific row in GPT for Excel, enabling you to test results on targeted content in your workbook. In the example below, GPT for Excel will run the next 5 empty rows starting from row 4.
The preview feature in GPT for Excel now works correctly when the previewed cell contains only "0".
The Custom prompt bulk tool preview now works correctly when you switch between different sheets in your workbook.
We've resolved issues with the function counts:
Count of active functions for users running their first function no longer shows -1.
Counts of all function statuses on sheets with empty A1 cells are no longer stuck to 0.
You can now stop bulk tool executions running in another sheet or spreadsheet.
We've resolved the issue where executions were incorrectly marked as "failed" in case of empty balance. Now, a message indicates that you need to purchase a pack.
You can now use your OpenAI Assistants directly in GPT for Sheets! Set up your OpenAI API key to make your Assistants available in the model switcher.
We've fixed an issue where selecting Show sources in the Search the web tool caused an incorrect execution count.
Claude 3.5 Sonnet is now available with an Anthropic API key. This model operates at twice the speed of Claude 3 Opus, and costs the same as Claude 3 Sonnet: $3 per million input tokens and $15 per million output tokens, with a 200K token context window.
Claude 3 Sonnet is no longer supported and GPT for Sheets formulas using this model will fall back to Claude 3.5 Sonnet. If you notice differences in responses, please contact support.
The Retry errors and Clear errors cells buttons now display after running a bulk tool when at least one row fails, rather than only when all rows fail.
We've fixed an issue where error counts in bulk tools were inflated for executions under 21 rows, especially with the gemini-1.5-flash model.
The execution button in bulk tools now shows Run all rows instead of the row count when switched to All rows, optimizing PC resources.
The Custom prompt bulk tool now features live prompt previews directly under the Prompt field, allowing real-time updates with the data in your sheet.
Bulk tools are now available for all users in GPT for Sheets. These robust tools allow you to Translate, Extract entities, or run a Custom prompt on a whole column at once, and are much easier to use than GPT functions.
Support for Google Gemini models is now available in GPT for Sheets. You can use gemini-1.5-flash and gemini-1.5-pro with your own Gemini API key to benefit from fast responses and very large context windows (1M tokens).
We've resolved an issue where the message "Interrupted!" was shown instead of "Column is full" when an empty row was found at the beginning of the sheet.
The Extract entities bulk tool is now available on GPT for Sheets. This tool allows you to automatically identify and extract key information, such as email addresses, names, and locations, from an entire column of text.
Reminder: Bulk tools are progressively being rolled out and are currently available to 50% of users, including the Translate and Custom Prompt tools.
We've fixed an issue in the Custom prompt bulk tool where the Run rows button was inactive and displayed Run 0 rows if the output column was selected before typing your custom prompt.
The GPT_WEB function and the Search the web bulk tool now deliver more accurate results, as we improved how GPT for Sheets and GPT for Excel prompt the Perplexity AI model.
We've fixed an issue where selecting All rows without specifying an output column caused an infinite loading screen, instead of a grayed out button, which could be mistaken for a system error.
The Search the web bulk tool now displays by default the source URLs in the column to the right of the search results. Sources will appear if this column is empty, allowing you to assess their reliability.
Executions now quickly stop when you cancel them, ensuring your balance is not affected. This change is especially effective for handling over 10,000 cancellations, with noticeable reductions in delays.
We've fixed an issue in the =GPT_FILL function that caused internal instructions to be included in the output.
When using Bulk tools, running small batches of up to 20 rows now completes in under 2 seconds, previously taking up to 10 seconds.
We've resolved an issue causing incorrect cached results to be returned when using GPT_VISION.
We've resolved an issue where using Bulk tools in Excel 2021 could incorrectly trigger a column full error when selecting empty output columns.
You can now preview your prompts even when you select a column that is filled with data in the Put the results in column dropdown of the Custom bulk tool.
We have just launched the Summarize bulk tool in GPT for Excel.
This tool allows you to summarize the content from each cell of a given column in a specified format. To access it, simply click on the Bulk tools tab in the GPT for Excel sidebar.
For instance, you can condense articles for a daily press review of the technology sector into concise 100-word summaries.
We've fixed a bug in the Bulk tools that causes interruptions and slowdowns when your spreadsheet contains large numbers of rows (above 45K rows).
We've resolved an issue where opening the Token counter page while selecting an empty sheet would cause an error.
GPT_VISION now uses OpenAI's latest model, GPT-4o, for image analysis. Additionally, you can now use GPT_VISION without an OpenAI API key.
You can now hide the instructions and shortcuts to clear the sidebar, and display them again whenever needed.
You can now log in using Firefox even with Enhanced Tracking Protection enabled.
GPT for Work extensions no longer support Azure models. These models were slower and not updated as quickly as OpenAI models. This change simplifies model selection.
Formulas using Azure models in their parameters fall back to the corresponding OpenAI models, with no change in costs.
This release fixes issues in the progress screen for bulk tools:
The total number of rows to be processed now displays correctly as soon as the execution starts.
Stopping the process no longer displays the remaining unexecuted rows as errors.
We have just launched the Translate bulk tool in GPT for Sheets.
This tool allows you to translate the content of an entire column, customize your translations with instructions, and use a glossary to guide the AI in the process. To access it, simply click on the Bulk tools tab in GPT for Sheets sidebar.
GPT for Sheets bulk tools are being rolled out progressively and are currently available to 25% of users.
This release includes the following fixes for the bulk tools:
Clicking on a bulk tool now navigates to the intended page, resolving the issue that sometimes redirected users back to the main bulk tools page.
Disabling the Exclude column headers option now works correctly, even if the Custom prompt includes a full column name reference, like {{ColumnName}}
.
Correcting the prompt in the Custom prompt tool now immediately removes the error message.
We fixed an issue where the Go to checkout button remained disabled when you tried to buy a new pack. Previously, you had to change the country to activate the button.
The Custom prompt bulk tool no longer crashes when the active sheet contains special characters in its header row.
Non-Latin characters now work correctly on the checkout page for Excel and Word, allowing all users to buy packs.
We are excited to announce that OpenAI's latest model GPT-4o is now supported by all our extensions for Microsoft Excel, Word, Google Sheets, and Docs.
GPT-4o matches GPT-4 Turbo's performance on text in English while being much faster and 50% cheaper in API costs. It is now the default model for all new users.
We have fixed bulk tool execution statuses to prevent indefinite loading when errors occur. Errors now correctly add to the error count, and bulk tools do not get stuck.
The bulk tools are being rolled out progressively and are currently available to 25% of users.
An error that displayed in the Custom prompt bulk tool when navigating between sidebar tabs or reopening the add-on has been fixed. This error did not impact functionality.
Web searches are now conducted with Perplexity AI's model llama-3-sonar-small-32k-online.
This update applies to the Web browsing bulk tool in GPT for Excel and to the GPT_WEB function in both extensions. Previously, the sonar-small-online model was used.
GPT for Excel Word can now use your Microsoft Teams authentication to sign you in.
We've updated the bulk tools to enhance user experience. Now, if you launch a prompt on multiple cells and your balance is too low, the system alerts you before any action takes place. You have the option to either run the prompt on as many cells as your remaining balance allows, or to add to your balance before proceeding.
On April 29th, our backend infrastructure started having an issue that resulted in many timeout errors for heavy users of GPT for Sheets. This issue is now fixed as of May 3rd 10am UTC.
In GPT for Sheets bulk tools, you can now:
Reference cells from different sheets
Select the number of cells you want to run
Free trial users can now use the GPT_WEB function to search the web for up-to-date information.
We have extended the cache duration in GPT for Sheets. Free users now benefit from a 7-day cache, and paid users enjoy a 30-day cache. The previous setting allowed a 6-hour cache for all users.
This change prevents unwanted Google Sheets refreshes of GPT formula results, and avoids the related costs.
We resolved an authentication issue that was preventing users who had never activated the Safe mode to access bulk tools.
Formula counts now automatically update after you Replace formulas, Retry errors, or Regenerate results.
Fixed range references from another sheet are now correctly highlighted in the Prompt field and accurately replaced in the Preview of the Custom bulk tool.
We are currently rolling out the first bulk tool for GPT for Sheets in Beta. It is now available to 25% of users.
This tool allows you to apply any prompt to an entire column and specify the destination column for the results, without writing a single formula. For instance, you can easily extract email addresses from column A and have the results appear in column B.
To access it, simply click on the Bulk tools tab in GPT for Sheets sidebar.
Replacing formulas with their results is now much quicker. All formulas are now replaced with their values when they are within the selected range or sheet, regardless of their nature or status.
Now, when you select all cells using cmd/ctrl+a, and click Regenerate selection or Replace formulas in selection, all formulas will update correctly.
Each cell now displays an estimated time of arrival (ETA) during Safe mode executions.
The token count now refreshes only when you press the Refresh button or expand the Token counter section, giving you control over when the data refreshes.
The latest OpenAI model (gpt-4-turbo-2024-04-09) is now used for gpt-4-turbo. Learn more
GPT_VISION now uses this gpt-4-turbo snapshot as per OpenAI latest recommendation, and gpt-4-1106-vision-preview and gpt-4-vision-preview are no longer supported.
You can now use the Mistral Large model from GPT for Sheets!
Simply enter a Mistral API key in the GPT for Sheets add-on and start using Mistral AI's best commercial model so far:
This model is currently free to use in GPT for Sheets, excluding API costs which are paid directly to Mistral API.
This release improves GPT_MATCH performance. Previously, executing several GPT_MATCH functions near their row limit (1300 rows) was noticeably slow. Now, these executions are up to 100 times faster.
The add-in now loads successfully, even on slow 3G networks.
This release fixes an issue where reaching the OpenAI API key rate limit would trigger an error in GPT for Sheets sidebar.
Your balance now remains visible in the footer when using an Anthropic model. Anthropic Claude 3 models are currently free to use, which means that executions run with an Anthropic model will not impact your balance.
The sidebar token counter remains visible regardless of the model selected.
When the cut-off limit is reached (max_tokens) while using GPT_WEB in safe mode, truncated formula results are correctly prefixed with a message to help you identify the issue and fix it by modifying the value of the Max tokens setting.
Selecting a model that didn't require an API key when using GPT_VISION, previously caused an error. Now, the system automatically defaults to the ‘gpt-4-1106-vision-preview' model to prevent these errors, regardless of the model initially chosen.
We've updated the model switcher. See the list of supported models.
You can access previous versions of OpenAI gpt-3.5-turbo and gpt-4-turbo models via the settings or function parameters.
We've updated the model switcher. See the list of supported models.
You can access previous versions of OpenAI gpt-3.5-turbo and gpt-4-turbo models via the settings or function parameters.
The correct default model gpt-3.5-turbo is now used, replacing the deprecated model gpt-3.5-turbo-16k.
Renaming your sheet when using bulk tools no longer throws errors.
Dragging formulas to up to 10,000 rows now runs smoothly without freezes.
GPT for Sheets prices have been updated to reflect the new GPT for Excel and Word prices, and the cost for input and output tokens is now also identical.
GPT for Excel and Word prices have dropped for all models:
gpt-3.5 turbo models cost $2.5 for 1M tokens, instead of $25 previously
gpt-4-turbo models cost $35 for 1M tokens, instead of $55 previously
gpt-4 models cost $65 for 1M tokens, instead of $85 previously
text-embedding-ada-002 costs $2.5 for 1M tokens, instead of $25 previously
Perplexity model costs $15 for 1K executions, instead of $50 previously
With an API key (excluding OpenAI API cost which you pay directly to OpenAI):
gpt-3.5 turbo models cost $1 for 1M tokens, instead of $15 previously
gpt-4 and gpt-4-turbo models cost $5 for 1M tokens, instead of $15 previously
text-embedding-ada-002 costs $1 for 1M tokens, instead of $15 previously
Packs purchased before March 13th, 2024 keep their original pricing.
We will continue to iterate on the pricing, please do not hesitate to send feedback, good or bad.
Running multiple requests simultaneously with the GPT_HSPLIT function no longer throws an error message.
You can now correctly insert the GPT_MATCH function example from the List of GPT functions into your sheet.
Double-clicking on the navigation arrow no longer results in a blank page or a login page with non-functional buttons.
You can now click on your balance to see the Google account currently used in GPT for Sheets. This provides a quick way to check if it matches the account used for pack purchases.
When you enable safe mode in GPT for Sheets, the Auto-replace formulas option is now enabled by default. This significantly speeds up executions. You can still disable it if you'd rather keep all the formulas.
Resolved a problem causing a blank sidebar and stuck executions upon opening for some users. You no longer need to reload the add-in.
Eliminated an error that occurred when launching the add-in after a period of inactivity.
Fixed an issue where the GPT_EDIT function ignored the edit_instruction
parameter.
You can now click on your balance to see the Microsoft account currently used in GPT for Excel and Word. This provides a quick way to check if it matches the account used for pack purchases.
Users in China can now purchase packs in GPT for Excel and Word.
GPT for Word no longer loads indefinitely if launched with an invalid API key. Instead, an error message prompts you to enter a valid API key.
Anthropic Claude 3 models are now supported with an API key:
Opus: Powerful model
Sonnet: Hard-working model
Haiku: Light and fast model
All Claude 3 models support vision to understand and analyze images and 200,000 token context windows. They are currently free to use in GPT for Sheets and Docs, excluding API costs which are paid directly to Anthropic.
claude-instant-1 and claude-2 are no longer supported. Formulas using these models in their parameters fall back to claude-3-haiku. If responses from this new model differ, let us know by contacting Support.
Fine-tuned models based on babbage-002 and davinci-002 now correctly stop generating tokens before reaching the cut-off limit (max_tokens) when appropriate, addressing the previous issue of producing text up to the limit regardless of context.
When you install and launch the add-in, the Custom instructions field is now filled by default with: "I want you to act as a text assistant. Provide the answer without any introductory phrase.”
We've addressed issues related to the GPT_MATCH function:
You can now successfully insert examples into your workbook via the List of GPT functions from the sidebar
GPT_MATCH formulas are now sequentially executed, to minimize the risk of errors that was caused by an overlap of executions
GPT for Word launched in beta last September. After major upgrades, we're moving to a paid model.
Here's what's new:
Improved chat user experience
Document selection preview
Upgraded models to version 0125
Ability to use without an API key
The pricing is the same as GPT for Excel.
Packs purchased on Microsoft Word can be used on Microsoft Excel, and vice versa.
The pricing in GPT for Excel has been dramatically simplified based on your feedback. Here's what's new:
The cost of input and output tokens is now identical.
Without an API key (OpenAI API cost included):
gpt-3.5-turbo models now cost $25 per 1M tokens
gpt-4 models now cost $85 per 1M tokens
gpt-4-turbo models now cost $55 per 1M tokens
text-embedding-ada-002 now costs $25 for 1M tokens
With an API key, the pricing is the same across all models at $15 for 1M tokens (OpenAI API cost is not included as you pay it directly to OpenAI through your API key).
Those new prices only apply to packs purchased after the release of the new pricing. Packs purchased before continue to be consumed at the prices that were listed when they were bought…
We will continue to iterate on the pricing, please do not hesitate to send feedback, good or bad.
We have just launched the Web browsing tool in GPT for Excel.
This tool allows you to search the web to get up-to-date information. You can ask questions about events, facts, people, URLs…
To access it, simply click on Web browsing from the Bulk tools area in the GPT for Excel sidebar.
The price is $50 or €50 per 1000 executions.
Special characters are now supported in row headers when using Bulk tools.
The chat experience has been enhanced to allow the following actions during OpenAI response generation:
Scroll up the responses
Access the Copy/Insert/Replace buttons
The model switcher has changed to make it easier to choose a model based on:
The cost
The model main advantage
You can now also see all the details of a model by hovering on the model.
With the Bulk tools, you can now choose a specific number of cells to run.
You can now:
Preview the selected text of your document above the prompt field
Scroll back to the selected part of the document by clicking on the selection preview
We've improved the reliability of GPT for Sheets executions when the context is very large.
You can now buy $599 and $999 packs to support your high consumption. These packs allow for around 799 and 1332 million input or output tokens with gpt-3.5-turbo, respectively. Like all packs now, they are valid for a full year.
OpenAI has upgraded the model under gpt-3.5-turbo to gpt-3.5-turbo-0125.
This model offers:
Higher accuracy at responding in requested formats
Lower pricing
Larger context window
We have just launched two new Bulk tools in GPT for Excel, that allow you to:
Classify the content of a column into predefined categories. For example, classify your user feedback into relevant categories for your Support team.
Extract relevant information from a column. For example, extract email addresses from text, regardless of their original format.
To access these and more tools, click on Bulk tools in GPT for Excel sidebar.
We solved an issue that prevented the selection of long AI responses for copying, replacing, or inserting text in your document.
Packs bought from now on expire one year after their purchase date.
Packs bought before this update still expire after 90 days.
A new free extraction tool, capable of processing up to three texts simultaneously, each up to 200 characters, is now freely accessible on the GPT for Work website.
With the Custom Bulk tool, you can now preview your prompts before running them.
The Copy button in the Translate tool now copies the translated text as expected.
The count of active functions now also includes paused functions. Previously, they were included as errors.
You can now buy $599 and $999 packs to support your high consumption. These packs allow for around 66 and 110 million input or output words with gpt-3.5-turbo in GPT for Sheets.
A new classification tool is available for free on the GPT for Work website. You can classify texts up to 200 characters into up to 3 predefined categories.
We are currently improving the onboarding process for GPT for Excel. An introductory video is now available to help all new users get started.
If your API key is valid, clicking the I've fixed the problem, retry button in the OpenAI API key not working error message will now effectively recognize and validate it, enabling access to GPT for Word.
We have just launched the Translate tool in GPT for Excel.
This tool allows you to translate an entire column and specify the destination column for the results. For instance, you can translate a whole column containing multiple languages in column A into a single, common language in column B. You can also customize your translations adding Translation instructions, and use the Translation glossary to guide the AI in the process.
To access it, simply click on the Bulk tools tab in GPT for Excel sidebar.
If a payment succeeds but there's a delay in the pack creation due to our billing service timing out with Stripe, you can now use a Refresh button to create the pack.
If you experience issues with the payment process due to network connection problems, a troubleshooting message will appear, providing instructions to assist you in resolving the issue.
You can now Copy part of the AI response or Insert it into your Word document.
A new translation tool is available for free on the GPT for Work website. You can translate up to 200 characters, customize your translations using the Glossary of special words, and add Custom instructions to guide the AI in the process.
This release expands the range of OpenAI models accessible in GPT for Sheets, with and without an API key:
gpt-4-turbo-preview (gpt-4-turbo)
gpt-4-0125-preview
We've updated the model switcher with the latest, most stable and used models. See our model comparator.
We clarified in the Settings section that Safe mode is only for gpt-4, gpt-4-turbo and GPT_WEB.
You can now access the Cost calculator from the Token counter section of the GPT for Sheets and GPT for Excel sidebar.
When you click on the Cost calculator button, the input tokens value is auto-filled into the Usage cost calculator, allowing you to see the cost implications based on the model selected.
You can now set Custom instructions to guide the AI with specific preferences and roles. These custom instructions are active until you choose to Reset to default. You can access this feature from the Settings icon.
The size of chat history sent to OpenAI is limited to a maximum of 4,000 tokens. OpenAI responses now take 2-3 seconds, no matter the actual size of the history. This new limit significantly reduces your costs.
Now, even with an empty balance, you can use the GPT for Excel sidebar to count the number of tokens in your cells, explore the list of functions, or view examples.
We've fixed an issue where sending chat messages was causing errors. This issue, caused by a minor update in the OpenAI API, affected responses from GPT-4, GPT-4-turbo, and GPT-3.5-turbo-16k.
To prevent further spending in any of your spreadsheets, you can now turn spending off. Auto-refreshes can still happen and cause data loss but will no longer incur costs. Replace formulas or use safe mode to avoid data loss.
This release fixes an error that incorrectly stated the document or selection was too long to process with some models like GPT-4. Now, you can process longer documents or selections up to the full model capacity.
The Usage cost calculator in our GPT for Excel and GPT for Sheets documentation now lets you compare pricing for two models and/or different languages. Click Add another model to compare to display two calculators side by side for easy comparison.
The model switcher now includes a price range indicator, showing the costs of different models at a glance. For detailed pricing, hover over the dollar signs.
The inline documentation for functions using specific models now includes the model name and the pricing details. This makes it easier for you to anticipate the costs of GPT_WEB, GPT_MATCH, and GPT_VISION function executions.
The Run button tooltip now clearly shows that exiting Pause will retrigger all formulas in #PAUSED! status. This helps you understand the cost implications before you click Run.
We've resolved an issue that led to timeouts for large requests, including your chat message, document selection, and chat history. You can now use the full capacity of models such as GPT-4 turbo, which allows 128K tokens.
The Custom prompt tool is now easier to use:
Tooltips now offer more details
Column references in the prompt field are highlighted for fast identification
The prompt field requires at least one column reference, reminding you to indicate the column your prompt applies to
The Run all button now displays the number of prompts that will be run
Safe mode is now available with GPT_WEB, allowing you to reduce timeouts and saving responses in the formula's last parameter. This prevents data loss and avoids extra costs during automatic Sheets refreshes.
We've updated the Model pricing comparator on the GPT for Work website. You can now toggle the switches to display the overall cost of function executions in GPT for Sheets and GPT for Excel, with or without an API key.
In this update, we've added a warning message for expired or revoked safe mode authorization. Just click Re-authorize safe mode to renew access.
Now, when facing issues with your OpenAI API key, a more detailed error message accompanies the Continue without an API key button.
Multi-line chat messages now disappear upon submission, making the AI response easier and faster to see.
We've addressed the Request too large error in safe mode, which previously occurred when a prompt combining multiple cells exceeded 50,000 characters.
The Pricing screen now provides clearer information. You can now easily determine the word capacity of a pack based on model choice and the language of your inputs and outputs.
The model switcher now displays tooltips for each model, indicating costs based on the model selected and the size of input and output tokens.
The new GPT_WEB function enables you to search the web to provide up-to-date information. It should be used whenever you ask questions about events, facts, people, companies for which freshness and accuracy is important.
It is built on perplexity.ai's online models.
The new GPT_WEB function enables you to search the web to provide up-to-date information. It should be used whenever you ask questions about events, facts, people, companies for which freshness and accuracy is important.
It is built on perplexity.ai's online model.
The Pricing screen now provides clearer information. You can now easily determine the word capacity of a pack based on model choice and the language of your inputs and outputs.
We have just launched the first tool for GPT for Excel.
This tool allows you to apply any prompt to an entire column and specify the destination column for the results, without writing a single formula. For instance, you can easily shorten product names from column A and have the results appear in column B.
To access it, simply click on the Tools tab in GPT for Excel sidebar. Watch this tutorial to get started.
You can now visualize how a specific piece of text is split into tokens and the overall count of words, characters and tokens, according to a specific model. Access this new tokenizer on our website.
GPT for Word now supports the gpt-3.5-turbo-16k model from OpenAI. This model lets you use context windows up to 16,000 tokens.
With this release, you can add your OpenAI organization ID in the GPT for Excel sidebar. If you are part of multiple organizations, it ensures that you get billed on the correct account.
We've fixed an issue that prevented GPT functions from executing in spreadsheets with Japanese, Danish, Chinese, or Korean language settings when GPT-4 safe mode was enabled.
We've resolved an issue that was blocking the execution of multiple functions within a single cell when the safe mode option was not enabled.
We now default to the Europe/Dublin time zone for date formatting when the system can't retrieve your spreadsheet's time zone, which eliminates errors.
We've resolved an issue that was preventing users with Japanese characters in their Microsoft account name from proceeding beyond the Sign in with Microsoft screen.
Your OpenAI API key may be valid but not working for different reasons, for example if you are out of OpenAI credits. In this case you will now see a dedicated screen that will help you solve the issue and then retry your API key.
Setting Max tokens to Auto is now only possible when an OpenAI chat model, such as gpt-3.5-turbo or gpt-4, is selected. Auto being specifically designed for OpenAI chat models, this prevents unnecessary selection with other models.
You can now regenerate formula results or retry errors when a single column or row is selected.
In this release, we've fixed an issue where you couldn't clear your OpenAI API key when prompted to do so.
Previously, if you had really long prompts, the favorite prompt button ⭐ would disappear from Prompt history. Now, no matter how long the prompt is, the favorite prompt button remains visible.
The Extension menu for GPT for Sheets now includes a new item: Retry errors. This provides a shortcut to open the sidebar, allowing you to manage errors and retry from there when you're done fixing.
We've fixed an issue where loading indicators were spinning for too long. Now, the loading indicators stop spinning upon action completion, whether it be counting formulas, replacing formulas with their results or regenerating results.
The new GPT_VISION function can analyze images and answer questions about them. You can use this function to generate product descriptions or captions for your images.
The redesigned sidebar in GPT for Excel now places the most useful buttons at the top, making them easier to access. You can now replace formulas with values, retry errors and pause executions in one click. You can also regenerate your results directly from the Settings screen after tweaking the parameters.
The authentication error in the GPT for Excel sidebar confirmation window has been resolved. Clicking Sign in with Microsoft now triggers a smooth sign-in process through silent authentication.
We've fixed an issue where GPT functions couldn't execute past column AZ when using GPT-4 safe mode.
GPT-4 safe mode is now available for all users, offering reduced timeouts and preventing loss of data and unexpected costs. It's a great option if you are using GPT-4 or GPT-4-turbo models.
This release expands the range of OpenAI models available from GPT for Excel:
gpt-4 and gpt-4-turbo models are now available without an API key.
The following models are now supported with an API key:
gpt-3.5-turbo-0301
gpt-3.5-turbo-0613
gpt-3.5-turbo-16k
gpt-4-0314
gpt-4-0613
gpt-4-32k
gpt-4-32k-0314
gpt-4-32k-0613
gpt-4-32k models are displayed only if you have access through your OpenAI key.
GPT for Excel now allows restarting the Microsoft sign-in process if you close the sign-in window, preventing endless loading.
gpt-4, and gpt-4-turbo (gpt-4-1106-preview) models are now also supported, alongside the gpt-3.5-turbo model (gpt-3.5-turbo-0613). You can switch between models from the model dropdown switcher.
We've resolved the following issues, which could affect spreadsheets using non-English language settings:
Clicking the Retry Errors button now triggers instant spinning, which eliminates the risk of multiple clicks.
Retrying errors previously disrupted some loading functions. Now, these functions operate smoothly.
The count of active functions accurately reflects both successful and loading functions.
Azure gpt-4 and Azure gpt-4-turbo and Azure gpt-3.5-turbo models are now supported by GPT for Excel without an API key. Learn more about Azure AI.
The model switcher dropdown that now clearly shows which models need an API key and which do not.
This release includes specific instructions for addressing NAME errors in spreadsheets. Learn more.
We've updated the system to refresh function counts before executing retries, ensuring the displayed numbers are always accurate and consistent. Also, the counter label is updated from "successful" to "active" for clarity.
The maximum number of requests per minute (RPM) in GPT-4 safe mode has doubled, increasing from 60 to 120.
If you hit OpenAI token per day limit (TPD), an error message informs you about the number of remaining tokens for the day. This helps you manage your usage more effectively.
The issue with the Replace formulas dropdown not responding to trackpads or long presses has been resolved.
In this release, we've corrected the token counter for the gpt-3.5-turbo-1106 model. Previously, it was showing an incorrect model limit.
The GPT_MATCH function is now available in GPT for Excel. It enables you to match the values of two columns by similarity. You can use it to match different spellings or even categories that are semantically related. This function uses the text-ada-002-embeddings model which is extremely fast and very cheap.
The value for Max input tokens now accurately reflects the full 128K token context window capacity of gpt-4-turbo (gpt-4-1106-preview).
The performance of GPT-4 safe mode has been improved, now delivering up to 60 requests per minute (RPM).
GPT functions are no longer stuck on “GPT Loading” during parallel executions across multiple spreadsheets when GPT-4 safe mode is enabled.
We've introduced the Tokens counter in the GPT for Excel sidebar. With this tool, you can get the number of tokens and input cost of the selected cell or range for the current model.
If you hit OpenAI token per day limit (TPD), an error message informs you about the number of remaining tokens for the day. This helps you manage your usage more effectively.
The redesigned sidebar in GPT for Sheets now places the most useful buttons at the top, making them easier to access. You can now replace formulas with values, retry errors and pause executions in one click. You can also regenerate your results directly from the Settings screen after tweaking the parameters.
Azure gpt-4 and Azure gpt-4-turbo models are now supported by GPT for Sheets. Learn more about Azure AI.
With this update, enabling GPT-4 safe mode automatically sets, by default, the Max tokens settings to Auto to make optimal use of your Tokens per minute quota. You can still specify a Fixed value for the Max tokens if needed.
The GPT for Work extensions now display a clearer error message with instructions if a problem occurs with an OpenAI payment.
With this release, you can add your OpenAI organization ID in the GPT for Sheets and GPT for Word sidebars. If you are part of multiple organizations, it ensures that you get billed on the correct account.
We resolved an issue with GPT_VISION not functioning when GPT-4 Safe mode was enabled and Max tokens setting was set to Auto.
This update ensures that if GPT functions get stuck in a #BUSY state, the GPT add-in sidebar now automatically opens, which resolves the issue.
OpenAI models GPT-4 and GPT-4 Turbo are now supported in GPT for Excel. Set up your own OpenAI API key in GPT for Excel to access these models and benefit from their broader general knowledge and advanced reasoning capabilities.
GPT formula caching now ignores sidebar settings. If the formula is unchanged, results are consistently retrieved from the cache.
Executions can now reschedule for a longer period of time. Previously, token expiration caused execution to fail after one hour.
The maximum number of requests per minute allowed when using GPT-4 safe mode has been increased from 30 to 60.
Truncated formula results are now prefixed with a message to help you identify the issue and fix it by modifying the value of the Max tokens setting.
We have fixed an error message that was displayed when hitting the tokens per minute rate limit. Previously, the message displayed the wrong model. This fix will help you better understand which model is causing the rate limit issue and allow you to take appropriate action.
We are excited to announce that we have renamed the add-in previously known as ChatGPT for Excel Word to GPT for Excel Word. This change was made to ensure consistency with our other extensions.
You can now find it in Microsoft AppSource under the name GPT for Excel Word.
In this release, we fixed an issue where the number of tokens for selected content was incorrectly estimated. It should have been zero, but was showing as equal to the number of tokens in your message. This was happening when the selection and the message were both large.
This was only an estimation error and did not affect the actual cost of token usage.
The new GPT_VISION function can analyze images and answer questions about them. You can use this function to generate product descriptions or captions for your images.
gpt-3.5-turbo-instruct is now available in the GPT for Sheets add-on. This model is specially designed to generate outputs that closely align with your instructions, ensuring more precise and relevant results.
When using GPT_EXTRACT, the output will no longer be split into multiple cells when the reference is a single cell.
Fine-tuned versions of the following models are now supported:
gpt-3.5-turbo
babbage-002
davinci-002
We are happy to announce that the Azure gpt-3.5-turbo model is now available, presenting a fast and efficient alternative to OpenAI's version. Also, you might prefer Azure OpenAI terms of service. Learn more about Azure AI.
This feature is in the process of being rolled out and is not yet available for all users.
The model provides faster response times compared to the OpenAI gpt-3.5-turbo, though it comes at a higher cost.
Using your own Azure API key to lower expenses is not yet possible.
As part of our enhancements for MacOS compatibility, we've improved the copying and pasting from the sidebar chat to accurately preserve line breaks.
This update resolves an issue where the first few characters of the responses were occasionally missing.
We've introduced a Tokens counter in the GPT for Sheets sidebar. With this tool, you can get the number of tokens and the associated cost based on your current model by selecting one or more cells.
This feature is in the process of being rolled out and is not yet available for all users.
This tool is not yet available for the newly supported models.
Your OpenAI API key may be valid but not working for different reasons, for example if you are out of OpenAI credits. In this case you will now see a dedicated screen that will help you solve the issue and then retry your API key.
The latest OpenAI model, GPT-4 Turbo, along with the updated version of GPT-3.5 Turbo (gpt-3.5-turbo-1106), are now accessible in GPT for Sheets and GPT for Docs.
GPT-4 Turbo is OpenAI's latest generation model. It's more capable, has an updated knowledge cutoff of April 2023 and introduces a 128k context window (the equivalent of 300 pages of text in a single prompt). The model is also 3X cheaper for input tokens and 2X cheaper for output tokens compared to the original GPT-4 model. The maximum number of output tokens for this model is 4096.
Updated version of GPT-3.5 Turbo now supports a 16K context window by default and improved instruction following. GPT-3.5 Turbo models are faster than GPT-4 and GPT-4 Turbo and return up to 4096 output tokens.
We've resolved an issue with the Auto-replace formulas option. Previously, it did not appear disabled when GPT-4 safe mode was turned off. Now, the option correctly reflects the status of GPT-4 safe mode, which reduces confusion.
The issue with GPT-4 safe mode activation not consistently saving with your spreadsheet has been resolved. This ensures your settings remain consistent across your work.
ChatGPT for Word now informs you when OpenAI servers are overloaded.
The latest OpenAI model GPT-4 Turbo (128k) is already available in our comparison tools:
We have also added gpt-3.5-turbo-1106 (updated version of GPT-3.5 Turbo) to the response time tracker and updated all pricing information to reflect the latest announcements.
The pricing for ChatGPT for Excel has changed. The price for 1 execution will now depend on its input and output size measured in tokens:
Small executions will be cheaper
Large executions will be more expensive
Save 50% by setting up your own OpenAI API key!
If you use your own OpenAI API key in ChatGPT for Excel, we will now bill you the same amount as OpenAI bills you. For example:
You spend $10 on OpenAI
We bill you $10
Your total cost is $20
In this release, the value of Max tokens can be set automatically for you, when you are using a GPT-4 model with the GPT-4 safe mode enabled. With the Auto option, Max tokens is automatically set according to the number of tokens in your input:
Max tokens = model token limit - input tokens
This allows you to benefit from the full token limit of GPT-4 without setting the value of Max tokens manually.
When launching ChatGPT for Word for the first time, you are now presented with a welcome screen and a video tutorial:
Watch the video to learn how to use the add-in.
Click Continue to set up the add-in.
GPT for Sheets now detects and ignores formulas that reference cells with errors also when GPT-4 safe mode is enabled. This means you won't face any unexpected delays or extra costs, as these issues are now identified upfront.
In this release, we've made the following changes:
GPT for Sheets settings are now saved at spreadsheet level. This means that you can configure distinct settings for each spreadsheet. When you open a new spreadsheet, it automatically uses your most recently changed settings. But, any changes you make to the settings after that will only apply to that specific spreadsheet.
Executing a function with an Anthropic model without an API key now displays a clear error message with instructions.
The maximum value for Max tokens is now set 200 tokens below the model token limit. For example, gpt-3.5-turbo has a capacity of 4096 tokens, so you can only set Max tokens up to 3896. This provides room for your input and prevents hitting the max token limit due to an excessively high Max tokens setting.
The Generate formula and Explain formula result fields no longer have a content overflow issue. You can now use the scroll bar to view long answers.
We improved the performance of GPT functions, reducing the risk of "Exceeded maximum execution time" errors.
We have added an LLM comparator on our website. This tool allows you to compare the most popular Generative AI LLM APIs based on tokens limit, price, rate limits, latency, language, and training data cutoff.
Error messages for Max tokens now display the maximum value for this setting in your current setup. These messages also recommend a different model if you need a higher model limit. This update makes it easier for you to locate and adjust the setting.
The new release saves chat history in your Word document. This means you can now retrieve your chat history even after reloading or closing the add-in or your Word document.
In this release, we've made the following changes:
Credits are now instantly displayed after purchasing a pack. There is no need to reload the add-in.
The VAT ID field is now validated. Invalid input is spotted, making it easier for you to identify and correct errors.
This release includes clearer messages for all OpenAI errors, making it easier for you to troubleshoot any issue you encounter.
Now, if the add-on encounters a rate limit error during API key validation, the system won't display an "Invalid API key" message. This change ensures that you only see this message if there's an actual authentication error.
GPT for Sheets launched in January, and quickly became a time-saving tool for early users. We've enhanced it significantly and are now moving to a paid model.
If you use your own OpenAI API key in GPT for Sheets, we will now bill you the same amount as OpenAI bills you.
For example:
You spend $10 on OpenAI
We bill you $10
Your total cost is $20
In the coming weeks, you can expect major improvements:
Much fewer timeouts and auto-refresh problems
Auto-replace formulas and other big UX improvements
This transition to a paid model is being rolled out progressively.
We've renamed GPT in Excel to ChatGPT for Excel Word. You can now find it in Microsoft AppSource under the name ChatGPT for Excel Word. After installation, the add-in unlocks the power of ChatGPT in Microsoft Word as well.
We are thrilled to announce that we released a new product: ChatGPT for Word is now available for all Microsoft users! 🎉
Free to use while in beta, you need to set your OpenAI API key and will pay the API cost directly to OpenAI.
With ChatGPT for Word, you can significantly enhance your Microsoft Word experience. Not only does it allow you to write, edit, and translate with greater efficiency, but it also leverages the content of your Word document for in-context results.
Features:
Chat with the chatbot within the add-in
Copy responses, insert them into your document, or replace some document content with them in a click
Select content in your document to use it as context for the next response
Using ChatGPT for Word will save you a lot of time by allowing you to:
Work on pieces of a long document
Structure your prompts in Word
Avoid constant copy/pasting in ChatGPT
Save ChatGPT answers in Microsoft Word for easier retrieval
Benefit from all the Microsoft Word editor tools, including real-time co-edition and change tracking
We've resolved a problem with the cache not functioning properly when switched off, then on again.
We've fixed an issue where the display of the creativity preset levels was delayed or incorrect.
In this release, we solved the following issues:
GPT_EXTRACT formulas now result in empty cells when nothing is extracted, instead of showing #REF errors.
You'll see a nicer error message if you attempt to enable GPT functions on a spreadsheet you did not create.
We have fixed a bug that could cause the add-in sidebar to crash when you were editing cells while GPT in Excel executed tasks.
We've introduced new model settings:
Custom instructions: Set specific preferences and roles for the model.
Top P: Control creativity by limiting the model to the most likely words or phrases with values like 0.1 for more focused results.
Frequency and Presence penalties: Reduce repetition with these OpenAI-specific settings.
The current model is now always included in the model status dropdown list at the top of the sidebar.
This release fixes an issue where the absence of the Anthropic version could lead to errors with Claude models.
We've modified GPT formula cache behavior to improve consistency and reduce unnecessary recalculations:
When caching is enabled in your spreadsheet, formula refresh is now restricted to error cases.
Caching is now formula-centric, and ignores sidebar settings. If the formula is unchanged, results are consistently retrieved from the cache.
When using a value from the cache, the formula result might therefore not reflect your latest sidebar settings.
This release includes the following improvements:
Model settings are now saved within the add-on, ensuring they are available when you reopen your document or use the add-on in a different document.
We have renamed and improved the ability to attach custom instructions to prompts. Previously called Set behavior and configured per prompt, this feature is now named Custom instructions and can be found in the Model settings section. With this update, custom instructions are set at the model level and are now saved within the add-on along all other model settings.
The Model list now shows only the models that are available to your currently set API keys.
When you select a Claude model, the default value for Max response tokens is now set to 10000. This change is designed to maximize the benefits of Anthropic models' larger token capacity when transitioning from a model with a smaller capacity.
text-ada-001, text-babbage-001 and text-curie-001 models, which will soon be retired by OpenAI, are demoted in GPT for Docs.
We have published a new template containing a comprehensive set of examples for major use cases. These examples are bundled into a spreadsheet that you can copy for Google Sheet
and into a downloadable workbook for Excel. Experience hands-on learning by copying and pasting the GPT formulas!
We have added the following functions to GPT in Excel:
GPT_SPLIT: Splits text semantically, e.g. by section, paragraph, customer. This function outputs all splits vertically in a column.
GPT_HSPLIT: Splits text semantically, e.g. by section, paragraph, customer. This function outputs all splits horizontally in a row.
GPT_EDIT: Applies edit instructions to the given text. The default task is to fix grammar and spelling.
We have also removed GPT_CONVERT from the list of available functions, because it was a duplicate of GPT_FORMAT. GPT_CONVERT remains usable by manually typing the function.
GPT_TAG and GPT_CLASSIFY now preserve the letter case of the tags and categories provided as input.
We have removed duplicates and unused functions from the list of available functions:
GPT_MAP: was identical to GPT_MATCH
GPT_CONVERT: was identical to GPT_FORMAT
GPT_CODE: was removed due to very low usage
GPT_MAP and GPT_CONVERT remain usable by manually typing the functions.
This release restores compatibility for formulas containing the deprecated max_tokens parameter. These old formulas now execute successfully, however any value specified for max_tokens will be ignored.
To configure this parameter, it is now recommended to use the sidebar setting Max response size.
GPT_TAG no longer returns an error if the tag list is not specified through the optional tags parameter. If the parameter is left blank, relevant tags are automatically suggested.
The model list now shows only the models that are available to your currently set API keys.
GPT for Sheets now takes into account locale settings for all countries in sidebar examples.
[insert] tags now work in documents containing tables with rows having varying column counts.
Claude models are now available in GPT for Docs!
Enter an Anthropic key in the GPT for Docs add-on and start using claude-instant-1 and claude-2 to benefit from:
Up to 100k tokens
Performance on par with OpenAI models
Affordable pricing: nearly 5 times cheaper than the 32K-context version of GPT-4
Price for 1 million tokens:
| | Claude 2 | GPT-4 32k | | -- | -- | | | Prompt | $11.02 | $60 | | Completion | $32.68 | $120 |
The formula replacement tool now displays the number of formulas replaced in near real time, updating the count every 7 seconds during processing.
If GPT_EXTRACT cannot find the value to extract, it will return blank instead of a REF error.
This update includes new formula examples on GPT in Excel home page and a refreshed design.
GPT_CLASSIFY doesn't hallucinate categories anymore and only returns results from the provided list of categories. If no category matches your input text, the function returns #N/A.
The execution rate limit has been reduced from 10 per second to 3 per second. This prevents reaching OpenAI rate limits and avoids errors.
The following functions were added to the GPT in Excel function family:
GPT_CLASSIFY classifies a given text into a single category.
GPT_CONVERT converts from one format to another. For instance json to table, or xml to json.
GPT_EXTRACT extracts data (like email addresses or company names) from a text.
GPT_FILL magically fills or cleans a range from a few examples.
GPT_FORMAT formats dates, currencies, addresses, names, etc. Fixes capitalization. And so much more.
GPT_SUMMARIZE summarizes text according to a given format (three sentences by default).
GPT_TABLE is like GPT, but outputs the results in a table. Very practical when the output is a table.
GPT_TAG applies user-defined tags to text or automatically suggests relevant tags if none are specified.
GPT_TRANSLATE translates text from one language into another. Works better if you specify the source language.
Dutch and Italian are now available as prompt languages.
With the latest update of our GPT for Work website, our pricing calculator and LLM response time tracker benefit from a redesigned user interface and integrate additional models.
We have simplified the pricing for GPT in Excel. Now, 1 credit = 1 execution. This is a change from the previous system where 1 to 8 credits could be used per execution.
You can no longer input out-of-range values in Creativity and Max tokens fields.
This release resolves an issue with the cache not functioning when the sidebar is closed. Now, the cache operates effectively regardless of the sidebar status.
This version includes the following improvements:
Gpt-3.5-turbo-0613: GPT in Excel now operates with this newly released model from OpenAI.
Reduced latency due to improved performance of the Microsoft authenticator.
Clearer display of your remaining credits.
GPT for Sheets and Docs now supports the gpt-3.5-turbo-16k model from OpenAI. This lets you use context windows up to 16,000 tokens, a big step up from the previous limit of 4,096 tokens.
We are excited to announce that our GPT in Excel add-in is transitioning to a paid model, effective immediately. You don't need to create and set up your own OpenAI API key anymore. Instead, buy credit packs that last 90 days. Each execution of GPT functions then costs credits.
More functions are also coming soon, stay tuned!
gpt-3.5-turbo is now the default model in GPT for Sheets and Docs.
The response time of gpt-3.5-turbo has significantly decreased which makes it a better trade-off between reliability, quality, and cost than text-davinci-003.
The Replace GPT formulas with result feature now handles large sheets with 50,000 cells or more to replace.
Error messages are more specific and provide clearer guidance
The Replace GPT formulas with results feature now disregards cells containing GPT formulas that are in error, loading, or paused states. This ensures that only valid formulas are processed.
GPT_MAP can only handle a maximum of 1300 rows due to Google's limitations. To aid in this, we've included a clear error message. If you encounter this message, you can resolve the issue by executing GPT_MAP in smaller batches to stay within the limit.
We've added the gpt-4-0314 model to our GPT for Docs extension!
In this new version of GPT for Sheets:
We've removed the formula replacement and cache management items from the Extensions menu. These features remain accessible from the GPT for Sheets sidebar.
We now only retrieve spreadsheet information when the GPT formulas controls section is open. This reduces unnecessary data usage.
We've updated the URL for sidebar assets in GPT for Sheets to fix blank sidebar issues.
We just launched a new version of the GPT for Sheets add-on that brings back the Set API key option in the Extensions menu.
OpenAI has reduced the RPM (Requests Per Minute) limit for free trial users from 20 to 3. This change could lead to errors. We have identified this specific error and updated the error message to recommend an upgrade.
The sidebar now opens automatically when you install the add-on.
We have added the following functions to GPT for Sheets:
GPT_SPLIT: Splits text semantically, e.g. by section, paragraph, customer. This function outputs all splits vertically in a column.
GPT_HSPLIT: Splits text semantically, e.g. by section, paragraph, customer. This function outputs all splits horizontally in a row.
You can now choose a model directly from the Selected Model Status list, enabling you to view the current response times of the models to guide your choice.
To ensure optimal performance, we've updated the default model for new users to text-davinci-003, following observed latency issues with the gpt-3.5-turbo model.
We've released a new version of the following functions:
GPT_TRANSLATE: we've introduced an [instructions] parameter for you to provide specific guidelines for your translations.
GPT_TAG and GPT_CLASSIFY: We also added an [example] parameter for you to provide a table of examples with two columns: the first column should contain the inputs and the second column should contain the outputs.
With this release, you can now:
Mark a prompt as favorite to save it permanently, ensuring it remains accessible even after you close the document or the add-on.
Select the language in which you write your prompts and custom instructions to optimize the model's response.
We solved an issue that was causing quotes to appear in the output.
A new page is now available on the GPT for Work website allowing you to track OpenAI and Azure OpenAI APIs response times.
The GPT_FILL function now neatly displays the output in a single cell when provided with a single-cell input.
We are pleased to announce that GPT for Sheets now supports GPT-4, just 24 hours after OpenAI's official release of this model. Enjoy the latest advancements in AI within your Google Sheets environment!
The GPT_MAP function now includes the top_k parameter. This parameter lets you specify the number of matching values to return per search key. By default, it returns only the best match.
We've introduced a new template showcasing how to use GPT functions in Google Sheets. This includes basic copywriting, data categorization, email address extraction, and more within your Google spreadsheet. Open the template and click Use template to duplicate it and test the GPT functions.
GPT_TABLE and GPT_EXTRACT now work better with table headers
GPT_CLASSIFY can now use the text-embedding-ada-002 model for classification, which will prevent GPT from inventing unwanted categories. To use it, simply specify the model in the function parameter
GPT_TAG has a new top_k parameter which allows you to control the maximum number of tags to be returned
Quotation marks are better handled so that they are not trimmed when they're wanted
Errors of authorization linked to multiple account should be better handled
We've made improvements to GPT_FILL, GPT_TABLE and GPT_EXTRACT
OpenAI released the new chatgpt gpt-3.5-turbo model last night, and it is now the default model in in all functions of GPT for Sheets and Docs. Enjoy 10x cheaper costs, faster and better quality responses.
We turbo released this, so if it breaks anything, (1) we're sorry and (2) please let us know here
Specifying tags in GPT_TAG is now optional: GPT_TAG will come up with tags on its own
GPT_TAG and GPT_CLASSIFY output in lowercase if categories are lowercase
We now automatically detect fine-tuned model most frequent separators
A few other details that make the overall experience better
GPT_MAP which enables you to map the values of two columns to find the best matches. You can use it to map different spellings or even categories that are semantically related. This function uses the text-ada-002-embeddings model which is extremely fast and very cheap
GPT_TREE which is more of an experimental function to generate trees of content in your spreadsheet.
We've also added a GTP function for those who find the letter T faster than the letter P! 😊
You can now use your fine-tuned models in GPT for Sheets! Simply specify your model in the model parameter of your GPT function.
We will automatically recognize or add the stop sequence of your prompt.
Datetimes are now better handled. (the devil really is in the details)
We now better handle incorrect values in all parameters
GPT_CONVERT now works better
The default max_tokens parameter has been increased to 500 in order to avoid to many cut responses
The devil is in the details!
ChatGPT outputs pure text by default. When this text contains numbers or dates, we want them as numbers or date objects in our spreadsheets in order to be able to operate on them.
GPT for Sheets now correctly recognises numbers and dates before writing them to your spreadsheet. This makes it much more usable to work with tables of data.
We have improved parsing of input in the prompts sent to OpenAI which will result in better quality output.
We have also improved our preset prompts in GPT for Docs for better output quality.
We've added an examples template spreadsheet directly accessible from the add-on menu.
Replacing formulas now works on all GPT functions.
We've made error messages easier to understand which should help you understand what to do when you input a parameter incorrectly, hit a rate limit, expire your free trial or when OpenAI servers are unavailable.
Woohoo! We just released a bunch of new functions that will make it much easier to make some frequent operations without writing any prompt:
GPT_TAG to apply tags. Simple and easy
GPT_CODE to generate code in your spreadsheet. Yes some people want to do this!
GPT_EDIT to edit text. By default it fixes spelling and grammar
GPT_EXTRACT to extract things like email addresses or companies from a paragraph. Works remarkably well
GPT_CONVERT to convert cells and ranges from one format to another: html, json, json-ld, nl-json, tables, csv, markdown. Incredibly powerful
GPT_SUMMARIZE to quickly summarize some text
GPT_TRANSLATE to quickly translate some text
GPT_CLASSIFY to classify into a single category (different from GPT_TAG where multiple tags can apply)
and finally GPT_CREATE_PROMPT which will help you create a prompt from many different cells and ranges in your spreadsheet by just listing them one by one as arguments
Fix a bug when the value parameter of GPT_LIST was empty
Improvements to prompt and response parsing that make our functions more robust
Changed the default cache setting to disabled for new users, and automatically enable it after 100 executions
GPT_LIST now supports a range as a prompt. This makes it much easier to give write a good, specific, and detailed prompt.
Replace formula with values now works on GPT_FILL
We now check if you're the owner of the spreadsheet before you can set an API key
We have switched the order of parameters in all functions: max_tokens now comes before model, since:
tweaking the max_tokens parameter is actually useful and needs to be done frequently for longer-form content
there is very little reason to use another model than the default which is text-davinci-003.
OpenAI responses are correctly split in paragraphs when relevant
New tooltips have been added to the interface
The GPT function can now take range as prompt. This is a huge improvement, and makes it much easier to give write a good, specific, and detailed prompt.
We've released a new magical function: GPT_FILL ! This one is really incredible to see in action. All you need to do is feed it some example rows and it will fill a range intelligently. Super useful for cleaning lists, where the right prompt can be difficult to put together.
You can now use the add-on from Google Docs, which is much more suitable for long-form content
You can type your prompt directly in the document or in the prompt box
This update brings two cost management features:
Replacing formulas with values, from the Extensions menu: once you're satisfied with the answers of GPT, you should replace the formulas with values so that they never get recalculated.
Cache management: we now cache answers of the same prompt by default for a duration of roughly 6 hours. This means that every answer you get will be reused if you ask the same question again or the formula gets recalculated will help lower your OpenAI costs. You can of course disable the cache if you prefer to explore variations rapidly and do not want to reuse answers. The cache is per spreadsheet.
We've released a new function: GPT_LIST !
It makes it as simple as pie to get one item per cell when you want multiple items at once, like “5 taglines for an ice-cream shop”.
It all started with a simple GPT function. The little function that could.
The Replace GPT formulas with results feature now disregards cells containing GPT formulas that are in error, loading, or paused states. This ensures that only valid formulas are processed.