business

Best AI for Excel

The best AI tools for Excel as of August 2026 — Microsoft 365 Copilot Agent Mode, Claude for Excel, GPT for Excel, Shortcut, Grok for Excel and Gemini in Google Sheets, with SpreadsheetBench scores, real pricing and the COPILOT() retirement deadline.

Updated August 2026

Quick answer: There is no single best AI for Excel, because the three jobs people mean by “AI in Excel” have different winners. For general in-workbook work inside a Microsoft 365 tenant, Microsoft 365 Copilot Agent Mode is the default: it is the only option built into Excel itself, it has been generally available on Excel for Windows since January 2026, and it lets you switch between OpenAI and Anthropic models from the Copilot pane. For accuracy-critical financial modelling and model auditing, Claude for Excel is the strongest purpose-built spreadsheet product — the independent SpreadsheetBench 2 evaluation found that no commercial spreadsheet product beat a scaffolded model, and that among the products tested “Claude for Excel achieves the strongest performance”. For bulk row-by-row work across thousands of rows, GPT for Excel holds the highest independently evaluated add-in score on SpreadsheetBench V1-Verified at 92.5% (GPT for Work, 14 May 2026). The caveat that applies to all of them: on end-to-end business workflows, every agent tested on SpreadsheetBench 2 scored below 35%, and debugging scores sat near zero — so check the output before it reaches a decision.

One date belongs at the top of this page. Microsoft retires the =COPILOT() worksheet function on 14 September 2026 (The Register, 17 August 2026). If you have formulas that call it, they stop working in under three weeks from this update. The migration section below covers what replaces it.

This page covers spreadsheet-native AI only: tools that read, write and edit cells inside Excel or Google Sheets. If your question is broader — analysing a CSV, querying a warehouse, or building a BI dashboard — see Best AI for Data Analysis instead, which covers notebooks, Power BI, Julius and warehouse-native agents. Every figure below carries its source and its date; where a vendor’s own number conflicts with an independent one, both appear.

The state of AI in Excel: August 2026

Four things changed the picture in 2026, and none of them is “the models got smarter”.

1. Excel got a real agent, then a model picker. Agent Mode moved from Excel for the web to general availability on Excel for Windows in January 2026, with Mac following days later (Microsoft). Unlike Copilot Chat, which answers questions, Agent Mode edits the workbook: it writes formulas, builds pivot tables, generates charts and runs multi-step tasks, showing a plan for approval before executing. The same release added multi-model reasoning, letting Microsoft 365 Copilot and Microsoft 365 Premium users choose between OpenAI and Anthropic models rather than accepting whatever Microsoft routed to.

2. Every major lab now ships an Excel add-in. Anthropic’s Claude for Excel went from a 1,000-tester research preview in October 2025 to all Pro subscribers on 24 January 2026 to general availability on 7 May 2026. SpaceXAI shipped Grok for Excel on 20 July 2026 (SpaceXAI). The spreadsheet is now a contested surface rather than a Microsoft monopoly, and the add-in sidebar is where the fight happens.

3. Microsoft killed the AI formula and Google kept its own. =COPILOT() never left preview and will be gone on 14 September 2026; Microsoft’s guidance is to use the Copilot side pane instead. Google Sheets still ships =AI() (also callable as =Gemini()), which generates, summarises, categorises, runs sentiment analysis and pulls live answers from Google Search (Google). That is a genuine divergence: Google treats AI as a spreadsheet function, Microsoft has decided it is a chat pane.

4. The benchmarks got honest, and the numbers got worse. SpreadsheetBench 2, released in 2026 by Renmin University with Aptura.AI, AfterQuery and Shortcut.AI, tests end-to-end business workflows rather than single manipulations — 321 tasks averaging 11.8 sheets each, with debugging tasks averaging 593.5 modified cells and financial-modelling tasks over a thousand. The top score is 34.89%. The gap between “AI can write your VLOOKUP” and “AI can complete your three-statement model” is the whole story of 2026.

How the tools rank for Excel work

Ordered by how much serious spreadsheet work each one can carry, not by brand size. Scores in the table are drawn from the benchmark sections below; the three SpreadsheetBench tracks are not comparable with each other, so each entry names its track.

Rank 1 is a distribution verdict, not a quality verdict. Copilot Agent Mode leads because it is already licensed and already in the ribbon for most Excel users; on measured spreadsheet accuracy it is beaten by three of the tools below it.

The benchmarks: what SpreadsheetBench actually measures

SpreadsheetBench is the benchmark the industry converged on, and it is the one Microsoft, OpenAI and Anthropic have each used to report Excel performance. It matters because its tasks are real: 912 questions scraped from public Excel forums such as MrExcel and Chandoo, attached to the original workbooks, scored online-judge style — every output cell must match the golden file exactly, with no partial credit. It was a spotlight paper at NeurIPS 2024.

There are now three tracks, and mixing them produces nonsense. Read each number with its track attached.

Track one: V1-Full (912 tasks, mostly self-reported)

This is where the platform vendors publish. Results are submitted by the vendors themselves rather than run independently, which is the reason for the spread.

Rows 2 to 4 are as collated by Decide on 31 January 2026 from each vendor’s own reporting. Google’s 70.48% matches the top overall score published on the SpreadsheetBench V1 page, so Gemini in Sheets currently holds the highest figure on the full 912-task set. Treat all four as vendor-run: SpreadsheetBench marks results evaluated by external third parties as unverified, and none of Microsoft, OpenAI or Anthropic has submitted to the independently run Verified track.

Track two: V1-Verified (400 tasks, independently run)

In late 2025 the SpreadsheetBench authors worked with Shortcut’s Fundamental Research Labs to curate a 400-task Verified set, removing ambiguous instructions and tasks that could not be scored reliably. On this track the benchmark team runs the evaluation itself against your API — there is no way to cherry-pick.

Positions 1 and 2 are as reported on 14 May 2026 by GPT for Work; positions 3 to 6 are as reported on 31 January 2026 by Decide. The leaderboard updates in weekly batches, so check the live board before quoting a rank.

Two facts are worth carrying away. First, the independently verified frontier sits between roughly 82% and 94% — high, and much higher than the vendor-reported V1-Full numbers, because the Verified set strips out unscoreable tasks. Second, none of the entries above is Microsoft, OpenAI or Anthropic. The independently measured top of spreadsheet AI is currently occupied by specialists.

Track three: SpreadsheetBench 2 (321 tasks, end-to-end workflows)

This is the hard one, and the most useful for anyone deciding whether to trust an agent with a real model. Tasks are workflow-level rather than atomic: complete a financial model, audit and repair a broken workbook, build a specified pair of charts. Debugging tasks require hundreds of cell edits on average; financial modelling tasks require over a thousand.

Source: SpreadsheetBench verified runs, sampled 13 July 2026 via BenchmarkList. Self-reported launch-post figures place Kimi K3 at 34.8%, Claude Fable 5 at 34.7%, GPT-5.6 Sol at 32.4% and Claude Opus 4.8 at 31.55%; those have not been independently verified on this track.

The single most important finding for buyers sits in the paper’s Figure 3, which ran four commercial spreadsheet products — Kimi Sheet, GLM in Excel, Claude for Excel and ChatGPT for Excel — against scaffolded models on a 30-example subset: none of the products outperformed the scaffolded models, and among the products, Claude for Excel performed best — at 15.4% on that subset. That is an independent result, published by researchers with no stake in Anthropic, and it is the strongest evidence available for Claude for Excel’s accuracy positioning. It is also a low bar: the honest reading is “best of the products” and “behind a plain scaffolded model” at the same time.

The debug column is the number to stare at. The best agent in the world repairs a broken workbook correctly 12% of the time. If your use case is “find what’s wrong with this model”, AI is a second pair of eyes, not an auditor.

Why the three tracks disagree

The same product can look like a 92.5% performer or a 35% performer depending on which set you run. Three things drive that gap:

• Task shape. V1 tasks are single manipulations from a forum post. V2 tasks are multi-stage deliverables — the debugging tasks alone average 593.5 modified cells. Nothing about a 92.5% on the former predicts the latter.
• Who runs the evaluation. V1-Full is vendor-reported; V1-Verified is run by the benchmark team. The unverified numbers are the ones to discount.
• Scaffolding, not model. SpreadsheetBench 2’s scaffold comparison holds the model fixed and swaps the agent harness, and the harness moves the score. This mirrors what we found in Best AI for Coding: the wrapper often explains more variance than the model inside it.

Practical takeaway: use V1-Verified to compare add-ins against each other, use SpreadsheetBench 2 to calibrate how much you trust any of them on real work, and ignore V1-Full unless you are comparing vendor marketing.

Best AI tools for Excel compared (August 2026)

1. Microsoft 365 Copilot Agent Mode — best built-in option

Price: Microsoft 365 Copilot Business add-on from $18 per user per month paid yearly, promotional from a $21 list price; Microsoft 365 Copilot enterprise add-on $30 per user per month with annual commitment, on top of a qualifying base licence (Microsoft). Microsoft 365 Premium for individuals is reported by third-party pricing trackers at $19.99 per month with Copilot in the apps included; Microsoft does not list it on the Copilot business pricing page, so verify directly. Where it runs: Inside Excel on the web, Windows and Mac. No add-in to install. Models: A picker offering Auto plus OpenAI and Anthropic options; at the January 2026 desktop launch these were GPT-5.2 and Claude Opus 4.5, and the line-up has moved since.

Agent Mode is the only tool on this page that is part of Excel rather than bolted onto it. It plans before it acts — Copilot maps out its intended steps for you to review, approve or amend before it touches the workbook, which matters when the workbook is a financial model. It cleans datasets, builds pivot tables, generates charts, writes formulas and iterates until its own validation checks pass.

Why it wins its slot: zero procurement friction. If your organisation already pays for Microsoft 365 Copilot, Agent Mode is free at the margin, governed by your existing tenant policy, and needs no security review of a third-party add-in. Nothing else here can say that.

Limitations: it is slower than the add-ins and more prone to doing things you did not ask for. In an eleven-task head-to-head run by GPT for Work in March 2026, Copilot Agent Mode was fastest in zero of eleven tasks, took 4 minutes 46 seconds to check formulas that a rival did in 16 seconds, processed all tabs when asked to format one, and converted a range into an Excel table unprompted. On two bulk tasks it fell back to plain text formulas instead of generating AI content, and it stalled after 15 minutes on a 10,000-row job. That test was run and published by a direct competitor, so read the speed figures as directional rather than neutral — but the failure modes it describes match the general reporting.

Best for: organisations standardised on Microsoft 365 that want one governed tool for everyone.

2. Claude for Excel — best for financial modelling and auditing

Price: Included on paid Claude plans — Pro at $20 per month, plus Max, Team and Enterprise (Anthropic). The add-in itself is free from the Microsoft Marketplace. Where it runs: Add-in sidebar for Excel on Windows, Mac and the web, installed from the Microsoft Marketplace. Models: Anthropic’s current frontier line, including Claude Opus 5 and Claude Opus 4.8.

Claude for Excel reached general availability on 7 May 2026 alongside Claude for Word and PowerPoint (Anthropic), having opened to all Pro subscribers on 24 January 2026. It answers questions about a workbook with cell-level citations, updates assumptions while preserving formula dependencies, debugs errors to their root cause, and edits pivot tables, charts and conditional formatting in place.

Two 2026 additions changed what it is for. On 11 March 2026 Anthropic shipped shared context between Claude for Excel and Claude for PowerPoint, so a single session can pull comparable company financials from an open workbook, build a trading comps table, and drop the valuation summary into the deck without re-explaining the dataset (VentureBeat). The same release added Skills — saved, reusable workflows available to a whole organisation from the sidebar — with a preloaded set covering model auditing for formula errors, populating DCF and LBO templates, and cleaning messy data ranges. Enterprises can run the add-ins through a Claude account or through an existing LLM gateway to Amazon Bedrock, Google Cloud Vertex AI or Microsoft Foundry.

Why it wins its slot: it is the only spreadsheet product with an independent third-party result behind its accuracy claim. SpreadsheetBench 2’s product comparison found it the strongest of the commercial spreadsheet tools tested. Cell-level citations are the reason finance teams adopted it first — a number you can trace is a number you can sign off.

Limitations: it does not scale to bulk. In the March 2026 competitor benchmark, Claude in Excel hit rate limits and overloading, failed to finish a 100-row and a 1,000-row fill, and could not start a 10,000-row task at all. It was, however, the fastest tool tested on pivot table creation. Read it as a precision instrument, not a batch processor.

Best for: analysts, finance teams and anyone whose priority is that the numbers are right and checkable.

3. GPT for Excel — best for bulk work and the =COPILOT() replacement

Price: Free install with starting credits and no card. Then pay-as-you-go credit packs from $29, valid 12 months from the most recent purchase, or a subscription from $25 per month (GPT for Work). Where it runs: Add-in for Excel on Windows, macOS and the web; a Google Sheets version also exists. Models: ChatGPT and Claude models, selectable per formula via a model parameter or set as a workspace default.

GPT for Excel, built by Swiss software company Talarian, holds 92.5% on SpreadsheetBench V1-Verified — 370 of 400 tasks correct on the first attempt, running on Claude Opus 4.7 at medium reasoning. That is the highest score any Excel add-in has posted on the independently run track, and it is the single most credible accuracy figure on this page. The company published a line-by-line review of all 30 misses: 19 were defensible readings of ambiguous prompts and 11 were genuine errors, of which 7 flip to success on more than half of re-runs.

It has two distinct halves. The AI Agent handles chat-driven and bulk work at up to 1,000 rows per minute and 1 million rows per run, per the vendor’s published specification. The GPT functions are a family of 16 spreadsheet formulas — =GPT(), GPT_TRANSLATE, GPT_CLASSIFY, GPT_EXTRACT, GPT_VISION, GPT_WEB and more — that reference cells and ranges, recalculate when inputs change, and fill down a column like any other formula.

Why it wins its slot: it is the closest like-for-like replacement for the retiring =COPILOT() function, and it needs no Microsoft 365 Copilot licence to run. On speed, its SpreadsheetBench write-up reports an average of 24 seconds per task with a median of 20 seconds and 90% of runs under 45 seconds, at an average API cost of about $0.14 per task.

Limitations: the head-to-head benchmark showing it fastest in 10 of 11 tasks is the vendor’s own, run on its own machine against its own competitors — informative, not neutral. It is also another vendor and another security review, which is a real cost in a regulated organisation.

Best for: anyone processing thousands of rows, and anyone with =COPILOT() formulas to migrate before 14 September.

4. Shortcut — best for DCF, LBO and comps modelling

Price: Free tier with 3 tasks per day and no card required; Pro reported at $49 per month, Team at $199 per month. These figures come from third-party software directories rather than a published price page, so confirm with Shortcut before budgeting. Where it runs: Web app and Excel plug-in.

Shortcut is an Excel agent aimed squarely at investment banking and corporate finance workflows: building DCF, LBO and three-statement models from raw inputs or PDFs, running sensitivity analyses, and auditing existing formulas. It scored 86.00% on SpreadsheetBench V1-Verified as of the January 2026 snapshot.

It has an unusual credential: Shortcut.AI is a named contributor to SpreadsheetBench 2 itself, and its Fundamental Research Labs co-developed the V1-Verified set with the original authors. That cuts both ways. It signals genuine research depth in spreadsheet evaluation, and it means Shortcut’s own benchmark results should be read with the relationship in mind. The company’s marketing claim that it beats first-year McKinsey and Goldman Sachs analysts 89.1% of the time in blind testing is a vendor claim with no published independent replication.

Best for: finance professionals who build the same model shapes repeatedly and want the structure generated rather than the cells filled.

5. Gemini in Google Sheets — best if you are not actually in Excel

Price: Included with Google Workspace Business Standard and Plus, Enterprise Standard and Plus, the AI Expanded and AI Ultra add-ons, Google AI Pro for Education, and consumer Google AI Pro and Ultra. Where it runs: The Gemini side panel in Google Sheets, plus the =AI() cell function.

On 22 April 2026 Google shipped the ability to build and edit entire spreadsheets from natural language in the Sheets side panel: end-to-end creation from a single prompt, side-by-side editing of existing models, and pivot tables and complex formulas without manual configuration. It also handles optimisation problems using Google DeepMind and OR-Tools research. Rollout began 22 April for Rapid Release domains and 6 May for Scheduled Release domains. As of 26 August 2026 this feature is available in the United States in English only — the single most important qualifier on the Google option.

Separately, =AI() (or =Gemini()) puts a model in a cell: generate text, summarise, categorise, run sentiment analysis, or pull live answers from Google Search. Its documented limits are worth knowing before you build on it: responses are text only, the function cannot see the rest of your spreadsheet or your Drive unless you pass a range, you cannot undo or redo it, nested AI functions are unsupported, and only the first 350 selected AI cells generate per action.

Why it wins its slot: Google’s 70.48% on the full SpreadsheetBench set is the highest published figure on that track, and unlike Microsoft, Google has kept a formula-level AI primitive rather than retiring it. Per the launch note, promotional higher usage limits ran through 15 July 2026, after which per-user limits apply.

Best for: teams whose spreadsheets live in Google Sheets. See Best AI for Data Analysis for Google’s warehouse and notebook tools.

6. Grok for Excel — best free add-in

Price: The add-in itself is free from the Microsoft Marketplace. SpaceXAI describes it as “a free Microsoft 365 add-in”; third-party reporting states that running tasks requires a paid Grok plan such as SuperGrok, SuperGrok Heavy or Business. Verify against your own account before planning around it. Where it runs: Sidebar add-in; third-party reporting indicates support for Excel 2019 and later on Mac, plus Microsoft 365 and web versions. SpaceXAI’s own page does not list version requirements. Models: Grok 4.5 is reported as the default for the Office add-ins.

Launched 20 July 2026, Grok for Excel reads the range you select and answers questions about it, writes formulas from a description, runs scenarios, and drops charts into the sheet beside the data. Answers cite the cells they came from, matching Claude’s approach. Its differentiator is connectors: the add-in can pull context from recent emails or files in SharePoint and Google Drive into the workbook.

Why it wins its slot: it costs nothing to install, so it is a cheap second opinion alongside whatever you already use.

Limitations: it is the newest tool here and has no published SpreadsheetBench result, so its accuracy is unmeasured — data not available. Anyone evaluating it should also weigh the governance question of routing workbook contents to SpaceXAI, which is a different answer for a hobbyist than for a regulated firm.

Best for: individuals who want capable in-sheet AI without adding a subscription.

7. ChatGPT — best for questions about a workbook, not edits to it

Price: Free tier; Plus at $20 per month. Where it runs: Outside Excel. You upload the file or connect a source.

ChatGPT remains the most common way people use AI on spreadsheet data, and its Advanced Data Analysis mode writes and runs real Python on an uploaded workbook, which makes it strong for exploratory questions and for showing its working. OpenAI’s ChatGPT Agent self-reported 45.5% on SpreadsheetBench V1-Full.

The structural limitation is the one that defines this page: it is not in your workbook. Formatting, formula dependencies, pivot tables and conditional formatting are things it discusses rather than things it edits, and round-tripping a file breaks the audit trail. For analysis of tabular data more broadly, ChatGPT is covered in depth on Best AI for Data Analysis.

Best for: one-off questions about a file you are happy to upload.

8. Lightweight add-ins — best on a small budget

A long tail of add-ins covers narrower jobs at lower prices. None has published a SpreadsheetBench result, so treat capability claims as unverified.

Best for: individuals who need formula and VBA help rather than an agent, and small teams testing before committing budget.

Feature comparison: the full matrix

The COPILOT() function retires on 14 September 2026

This is the most consequential dated fact on the page, so it gets its own section.

Microsoft introduced =COPILOT() in August 2025 for Beta Channel users with a Microsoft 365 Copilot licence, later extending it to Excel for the web through the Frontier programme. It let you send an instruction to Copilot from a worksheet cell. It was scheduled for general availability in 2027. Instead, Microsoft updated its roadmap to say: “We have decided not to move forward with this feature.” Beginning 14 September 2026, the COPILOT function will no longer be available in Excel (Microsoft support; reported by The Register, 17 August 2026).

Microsoft’s guidance is to use the Copilot side pane, which it says provides many of the same capabilities. That guidance replaces a formula with a chat window, and the two are not equivalent: a formula recalculates when its inputs change, drags down a column, and lives inside the workbook’s logic. A side-pane conversation does none of those things.

For the record, =COPILOT() was never a comfortable production tool even while it existed. Per Microsoft’s own support page, it required a paid Microsoft 365 Copilot add-on licence or Microsoft 365 Premium, was capped at 100 calculations every 10 minutes, saw only the cells passed as context arguments with no access to other workbook or enterprise data, and would not calculate in workbooks labelled Confidential or Highly Confidential. Microsoft’s launch post advised that its output “should be reviewed and validated for accuracy, especially for critical business decisions or reports”. Because it never left preview, it should not have been load-bearing in a production workflow — but if it is in yours, you have until 14 September.

What to do about it:

1. Find your exposure. Press Ctrl+H (Cmd+H on Mac), set Look in to Formulas, and search for =COPILOT( across your workbooks.
2. Choose a replacement. The nearest like-for-like is the GPT function family in GPT for Excel, where =COPILOT("prompt", A1) becomes =GPT("prompt", A1) with arguments carried over unchanged. Unlike the retiring function it exposes a model parameter, is not capped at 100 calls per 10 minutes, and does not require a Copilot licence.
3. Or move the work to an agent. If the formula was doing bulk enrichment, an agent sidebar handles it better than a filled-down formula ever did.
4. If you are in Google Sheets, do nothing. =AI() is not affected.

Use-case specific recommendations

For financial modelling and model auditing

Winner: Claude for Excel (included on paid Claude plans from $20 per month)

It is the only spreadsheet product an independent benchmark team singled out as best-in-class, cell-level citations make its answers checkable, and its preloaded Skills cover DCF and LBO template population and formula-error audits directly. Alternative: Shortcut, if you build the same model shapes repeatedly and want structure generated rather than cells filled. Caveat: the best agent on SpreadsheetBench 2 repairs a broken workbook correctly 12% of the time, so an AI audit finds errors — it does not certify their absence.

For bulk row-by-row processing

Winner: GPT for Excel (free credits, then from $25 per month)

It is the only tool on this page that documents throughput — up to 1,000 rows per minute and 1 million rows per run — and the only one that completed a 10,000-row generation task in the March 2026 comparison. Alternative: none of the built-in options; Copilot fell back to text formulas and Claude for Excel hit rate limits on the same tests.

For organisations standardised on Microsoft 365

Winner: Microsoft 365 Copilot Agent Mode ($18 to $30 per user per month depending on plan)

It is already in the ribbon, already inside your tenant’s governance boundary, and needs no third-party security review. The model picker lets you route heavy reasoning to Anthropic models without leaving the Microsoft interface. Alternative: add Claude for Excel for the finance team specifically; the two coexist fine.

For teams working in Google Sheets

Winner: Gemini in Google Sheets (included in Workspace Business Standard and above)

It holds the highest published score on the full SpreadsheetBench set at 70.48%, builds entire spreadsheets from one prompt, and keeps a formula-level =AI() primitive that Microsoft has abandoned. Caveat: spreadsheet building is United States English only as of 26 August 2026.

For learning Excel and writing better formulas

Winner: Ajelix or Numerous.ai (from free)

Formula and VBA generation from plain language is a solved, cheap problem, and you do not need an agent licence for it. Alternative: ChatGPT on the free tier explains what a formula does as well as anything paid.

For the lowest possible cost

Winner: Grok for Excel (free add-in)

A capable sidebar with cell citations and SharePoint and Google Drive connectors, at no installation cost. Caveat: third-party reporting says a paid Grok plan is needed to actually run tasks, and it has no published accuracy benchmark.

For privacy and compliance

Winner: Claude for Excel via an LLM gateway

It is the only option that can be routed through Amazon Bedrock, Google Cloud Vertex AI or Microsoft Foundry inside an existing compliance setup, rather than through a consumer account. Alternative: Copilot Agent Mode, which stays entirely inside your Microsoft 365 tenant boundary. Avoid: uploading workbooks with regulated data to a general chat tool.

For ad-hoc questions about a file

Winner: ChatGPT ($20 per month, capable free tier)

It writes and runs real Python on an uploaded workbook and shows its working, which is the right shape for exploration. Caveat: it cannot edit your live workbook, so anything you keep has to be moved back by hand.

Pricing comparison: what you’ll actually pay

All figures in USD, checked 26 August 2026. Add-on prices exclude the underlying Microsoft 365 or Google Workspace licence.

Sources for the table: Microsoft figures from the Microsoft 365 Copilot pricing page; GPT for Excel from its pricing page; Gemini availability from the Google Workspace launch note; Grok terms from SpaceXAI plus third-party reporting. Microsoft 365 Premium, Shortcut, Ajelix and Numerous.ai figures are as reported by third parties and are marked accordingly. Prices move; verify before committing budget.

The genuinely useful comparison is cost per unit of work, not cost per seat. GPT for Excel’s SpreadsheetBench run averaged about $0.14 per task at provider API pricing, with 90% of runs under $0.24 and a worst case of $1.40. No other vendor on this page publishes a per-task cost, which makes cross-tool cost comparison impossible today — data not available.

What Excel users actually report

Speed decides adoption more than accuracy does. The recurring complaint about Copilot Agent Mode is not that it is wrong, it is that a task taking one to five minutes (GPT for Work, March 2026) is slower than doing it yourself. An agent only saves time if it beats you to the answer; below that threshold people stop opening the pane.

Unrequested edits are the trust-killer. Agent Mode processing every tab when asked to format one, or silently converting a range into an Excel table, are the behaviours users cite when they turn it off. This is why plan-before-execute — now present in Copilot, Claude and Gemini — became the standard pattern in 2026.

Citations changed what finance teams will accept. Claude for Excel and Grok for Excel both cite the cells behind an answer. Users consistently report that a traceable number is the difference between an AI output they will put in front of a committee and one they will not.

Progress visibility matters on long jobs. On bulk tasks, the complaint about both Copilot and Claude in Excel was the absence of a progress indicator: users could not tell whether a job was running or had silently stalled. For anything that takes minutes, knowing where you stand is a feature.

Nobody uses the model picker. Independent reporting on Microsoft 365 Copilot’s model switcher, including Office Watch, notes it is rarely touched, and that the selection resets to the default each time a workbook is closed — a defaults problem rather than a capability problem, and worth knowing if you assumed your team was getting the model you chose.

Recent developments reshaping AI in Excel (2026)

• 24 January 2026 — Claude for Excel opens to all Pro subscribers. Anthropic ends the invite-only phase, adds multi-file drag and drop, stops the add-in overwriting existing cells, and supports longer sessions through automatic compression (The Decoder).
• 27 January 2026 — Agent Mode reaches general availability on Excel for desktop. Windows first, Mac days later, extending beyond the December 2025 Excel for the web release, and adding a multi-model picker across OpenAI and Anthropic models (Microsoft).
• 11 March 2026 — Claude gains shared context and Skills across Excel and PowerPoint. One session spans workbook and deck; saved workflows become one-click actions for a whole organisation; enterprises can route through Bedrock, Vertex AI or Microsoft Foundry (VentureBeat).
• 22 April 2026 — Gemini builds entire spreadsheets in Google Sheets. End-to-end creation from a single prompt, side-by-side editing, and optimisation problems solved using DeepMind and OR-Tools research, at a claimed 70.48% on the full SpreadsheetBench set. United States English only (Google).
• 7 May 2026 — Claude for Excel, Word and PowerPoint reach general availability. Anthropic completes the Office add-in suite (Anthropic).
• 14 May 2026 — GPT for Excel posts 92.5% on SpreadsheetBench V1-Verified. The highest independently evaluated score any Excel add-in has published, with a full public breakdown of every miss.
• July 2026 — SpreadsheetBench 2 results land. Every agent scores under 35% on end-to-end workflows; commercial spreadsheet products fail to beat scaffolded models; Claude for Excel is named the strongest product tested.
• 20 July 2026 — SpaceXAI ships Grok for Excel. A free Marketplace add-in with cell citations and SharePoint and Google Drive connectors, alongside Word and PowerPoint versions.
• 16–17 August 2026 — Microsoft confirms the =COPILOT() retirement. A retirement notice appears on Microsoft’s support page for the function and is reported publicly the next day. The function disappears on 14 September 2026, 13 months after its preview debut and having never reached general availability.
• August 2026 — Excel’s monthly update focuses on Copilot. Microsoft’s August release adds a change-history skill that summarises how a workbook has evolved and lets users undo specific edits or revert to an earlier state, per Neowin.

Frequently asked questions

What is the best AI for Excel in 2026?

There is no single winner, because the tools split by job. For general work inside a Microsoft 365 tenant, Microsoft 365 Copilot Agent Mode is the practical default, since it is built into Excel on the web, Windows and Mac and needs no extra add-in. For financial modelling and auditing, Claude for Excel is the strongest choice — SpreadsheetBench 2’s independent product comparison found it the best-performing commercial spreadsheet product tested. For bulk row-by-row processing across thousands of rows, GPT for Excel is the strongest, holding the highest independently evaluated add-in score at 92.5% on SpreadsheetBench V1-Verified.

Is Copilot or Claude better for Excel?

They win different things. Microsoft 365 Copilot Agent Mode is better on distribution and governance: it is native to Excel, covered by your existing tenant policy, and requires no third-party add-in review. Claude for Excel is better on measured accuracy and traceability: it gives cell-level citations, preserves formula dependencies when it edits, and is the only spreadsheet product an independent benchmark team named as strongest in its class. Many finance teams run both — Copilot as the organisation-wide default, Claude for Excel for the analysts who build the models.

Is the COPILOT function in Excel being removed?

Yes. Microsoft retires the =COPILOT() worksheet function on 14 September 2026. It was introduced in August 2025 as a preview feature for Beta Channel and Frontier users, was scheduled for general availability in 2027, and Microsoft has now stated it will not move forward with the feature. Microsoft’s guidance is to use the Copilot side pane instead. If you have workbooks containing =COPILOT() formulas, migrate them before that date — the nearest like-for-like replacement is the =GPT() function family in the GPT for Excel add-in, which maps argument-for-argument.

Can AI edit my Excel file directly, or only answer questions about it?

Both, depending on the tool. Microsoft 365 Copilot Agent Mode, Claude for Excel, GPT for Excel, Shortcut and Grok for Excel all edit cells, formulas, charts and formatting inside the open workbook. ChatGPT and other general chat tools work on an uploaded copy instead, which means anything you want to keep has to be moved back into your file by hand and the formula audit trail is broken in the process. If in-place editing matters, choose a native feature or a sidebar add-in rather than a chat window.

How accurate is AI in Excel?

Less accurate than the demonstrations suggest, and the gap widens with task complexity. On single manipulations from the SpreadsheetBench V1-Verified set, the best agents score between roughly 82% and 94%. On SpreadsheetBench 2’s end-to-end business workflows — completing a financial model, repairing a broken workbook — every agent tested scored below 35%, and on debugging tasks specifically the best model scored 12%. Treat AI output in a spreadsheet as a draft that needs checking, and prefer tools that cite the cells behind an answer.

What is the best free AI for Excel?

Grok for Excel is the strongest free option: the add-in installs at no cost from the Microsoft Marketplace, reads the range you select, writes formulas, builds charts and cites the cells behind its answers. Third-party reporting indicates a paid Grok plan is required to actually run tasks, so verify against your own account. For formula help specifically, Ajelix has a free tier of 3 messages per day and ChatGPT’s free tier explains and writes Excel formulas competently.

Do I need a Microsoft 365 Copilot licence to use AI in Excel?

No. A Microsoft 365 Copilot licence is required only for Microsoft’s own features — Copilot Chat and Agent Mode in Excel — which cost from $18 per user per month for organisations up to 300 users and $30 per user per month on enterprise plans, on top of a base Microsoft 365 licence. Every third-party add-in on this page, including Claude for Excel, GPT for Excel, Shortcut and Grok for Excel, runs in standard Microsoft 365 Excel with no Copilot licence at all.

Which AI is best for financial modelling in Excel?

Claude for Excel for auditing and editing existing models, Shortcut for building new ones. Claude for Excel ships preloaded Skills for populating DCF and LBO templates and auditing models for formula errors, cites the cells behind every answer, and was named the strongest commercial spreadsheet product in SpreadsheetBench 2’s independent comparison. Shortcut is purpose-built for DCF, LBO and comps construction and scored 86.00% on SpreadsheetBench V1-Verified. Neither is reliable enough to replace a review: the best agent measured repairs a broken workbook correctly only 12% of the time.

Does Google Sheets have an AI function like Excel’s COPILOT?

Yes, and unlike Microsoft’s, it is not being retired. Google Sheets offers =AI(), also callable as =Gemini(), which generates text, summarises, categorises, runs sentiment analysis and retrieves live information from Google Search. It requires an eligible Google Workspace or Google AI plan. Its documented limits are that responses are text only, it cannot see the rest of your spreadsheet unless you pass a range, it cannot be undone or nested inside other functions, and only the first 350 selected AI cells generate per action.

How much does AI for Excel cost?

Between nothing and $30 per user per month for the in-workbook tools. Grok for Excel’s add-in is free to install. Claude for Excel is included on Claude Pro at $20 per month. GPT for Excel starts with free credits, then $25 per month or credit packs from $29. Shortcut has a free tier of 3 tasks per day, with Pro reported at $49 per month. Microsoft 365 Copilot is $18 per user per month for organisations up to 300 users, promotional from a $21 list price, or $30 per user per month on enterprise plans, in both cases on top of a base Microsoft 365 licence. Gemini in Google Sheets is included with Google Workspace Business Standard and above at no extra charge.

Can AI write VBA macros and Excel formulas?

Yes, and this is the most reliably solved task on the list. Every tool on this page writes Excel formulas from a plain-language description. For VBA specifically, Ajelix generates VBA scripts and automation code for free on its entry tier, and general models handle it well — Claude and GPT models both write and explain VBA competently. If macros and scripts are your main need rather than in-sheet agents, the cheap tools are sufficient; see also Best AI for Coding for the broader picture on code generation.

How to choose in August 2026

Work through it in this order.

Start with where your data lives. If your spreadsheets are in Google Sheets, the decision is made: Gemini in Sheets, with the United States English caveat. Everything else on this page assumes Excel.

Then ask what you already pay for. If your organisation has Microsoft 365 Copilot licences, Agent Mode costs nothing extra and clears governance automatically. Start there and only add a second tool when it fails you at something specific.

Then match the tool to the job. Precision work on models goes to Claude for Excel. Volume work across thousands of rows goes to GPT for Excel. Building new financial models from templates goes to Shortcut. These are different products for different problems and stacking two of them is normal.

Then check the calendar. If any workbook you own contains =COPILOT(), that migration has a hard deadline of 14 September 2026 and should be handled before any tooling decision.

Finally, calibrate your trust. The most useful number on this page is not 92.5% — it is 12%, the rate at which the best available agent correctly repairs a broken workbook. AI in Excel is genuinely good at generating, formatting, filling and explaining. It is not yet good at guaranteeing that a model is right. Use it to do the work, and keep a human on the review.

Related reading: Best AI for Data Analysis for CSVs, warehouses and BI; Best AI for Business for the wider workplace stack; Best AI Models for the model layer underneath all of these tools.