Top 10 AI Agents for Excel Analysis, Ranked (August 2026)

---

title: "Top 10 AI Agents for Excel Analysis, Ranked (August 2026)" slug: "top-10-ai-agents-for-excel-analysis-ranked-2026" date: "2026-08-05" author: "O-mega Team" excerpt: "Benchmark-backed ranking of the best AI Excel agents in August 2026. Verified scores, live pricing, the Copilot rename, and the new model cycle."

The honest, benchmark-backed ranking of every AI agent that can actually analyze your spreadsheets in August 2026

Two of the biggest facts in this category changed names or generations in the last five months, and most rankings still have not noticed. Microsoft quietly renamed Excel's Agent Mode to Edit with Copilot around March 2026 and removed the OneDrive requirement, so it now works on local Excel files - Chris Menard Training. Then the model cycle turned in a single July fortnight: OpenAI shipped GPT-5.6 in three variants on July 9 - Engadget, and Anthropic shipped Claude Opus 5 on July 24 with a testimonial claiming 9 percentage points higher accuracy on hard financial-modeling tasks - Anthropic. A guide that still says "Agent Mode" and "GPT-5.5" is describing a market that no longer exists.

The platform war itself, meanwhile, is settled infrastructure: Microsoft, OpenAI, Anthropic, and Google all ship native agents inside the spreadsheet, and the interesting questions moved down a level. Which agent survives contact with a real workbook? The independent data is unusually good for an AI category: SpreadsheetBench exact-match scores, an analyst-graded financial modeling ranking from Wall Street Prep, an 11-task speed head-to-head, and a 15-tool accuracy test from AIMultiple. The spread those tests reveal is enormous: the best agent completes 92.5% of real-world spreadsheet tasks, Microsoft's own published figure for its Excel agent is 57.2%, and some tools that market themselves as "AI for Excel" score literally zero.

One honesty note before the ranking, because this category's search results have become a swamp. A wave of programmatic SEO pages now ranks for "best AI Excel agents" by placing the page's own product at #1 with self-run, unpublished benchmark claims that no third party can reproduce. This guide does the opposite: every score below names its source, every price was checked against a live pricing page on August 5, 2026, we did not run a private lab benchmark and we refuse to invent one, and where no independent test of a tool exists, the table says so in plain text. What you lose in false precision you gain in numbers you can actually verify.

Contents

1. What Changed by August 2026: The Rename and the New Model Cycle
2. The Benchmark Reality Check: What the Numbers Actually Say
3. The Top 10 AI Agents for Excel Analysis, In Depth
4. Head-to-Head: Edit with Copilot vs Claude vs ChatGPT vs Gemini
5. Models Under the Hood: Opus 5, GPT-5.6, and Why the Engine Matters
6. Accuracy, Failure Modes, and Human Oversight
7. Which Agent for Which Job: A Persona-Based Selection Framework
8. Deadpool and Rename Watch: Rows, Akkio, Copilot Pro, Agent Mode
9. Beyond the Spreadsheet: When Excel Analysis Is Just One Step
10. Conclusion: The Decision Framework

The Master Ranking

Four criteria, weighted by what actually determines success in spreadsheet analysis work. Verified accuracy (30%) counts only performance on named, independent benchmarks (SpreadsheetBench, Wall Street Prep, AIMultiple), never vendor claims. Workbook integration (25%) measures how natively the agent reads and edits real Excel files: direct cell edits, formula preservation, pivot and chart handling, protection against overwrites. Speed and scale (20%) measures throughput on large workbooks, where the only published head-to-head shows order-of-magnitude gaps. Value (25%) measures what you pay per seat for the capability you get, at prices we opened this week.

Read the scores for what they are: a compression of named third-party measurements into one sortable column, so you can scan ten tools in ten seconds. They are not the output of our own laboratory, and any cell resting on thin evidence says so inside the cell. That is the difference between this table and the self-scored tables now flooding this query's search results.

Reading the final column top to bottom: 9.3, 8.0, 7.5, 7.3, 7.0, 6.9, 6.3, 5.9, 5.8, 5.7. The order holds, and it still produces three uncomfortable results. The top spot belongs to a specialist vendor, not to Microsoft, OpenAI, or Anthropic, because GPT for Work posted the only independently documented 90%+ SpreadsheetBench score for a commercial agent. Edit with Copilot ranks sixth despite literally living inside Excel, because the published accuracy and speed data remain brutal. And Shortcut, a tool most Excel users have never heard of, outranks Microsoft's agent because an analyst-graded evaluation scored its financial models above every big-lab competitor - Wall Street Prep.

1. What Changed by August 2026: The Rename and the New Model Cycle

If you evaluated this category even three months ago, two structural updates invalidate parts of your notes. The first is naming and reach. What Microsoft launched in September 2025 as Agent Mode in Excel reached desktop general availability on January 27, 2026, complete with a multi-model switcher offering both OpenAI and Anthropic models inside a Microsoft product - Microsoft Tech Community. Around March 2026, Microsoft renamed the feature to Edit with Copilot across Excel, Word, and PowerPoint, and removed the biggest practical blocker: it now works on local files, not just workbooks stored in OneDrive or SharePoint - Chris Menard Training. Microsoft's own support documentation now describes the capability as an "Allow editing" mode that reshapes data, merges sheets, and builds multi-element reports directly in the workbook - Microsoft Support. If you search for "Agent Mode" today you are searching for a retired name; this guide uses the current one throughout, bridging to the old name only where readers arriving from older coverage need the connection.

The rename was not cosmetic housekeeping. The local-files change converts the feature from "works on the subset of your workbooks that live in the right cloud folder" to "works on the spreadsheet you actually have open," which was the single most common practical complaint in the preview period. By April 22, 2026, Microsoft declared the agentic capabilities in Word, Excel, and PowerPoint generally available, reporting Excel engagement up 67%, new-user retention up 50%, and satisfaction up 65% in the month after the desktop rollout - Microsoft. Those are vendor metrics and deserve the usual discount, but the direction matches what happens across the agent market whenever capability meets users inside the tool they already use.

The second structural update is the July model cycle, and it hit both frontier labs within fifteen days. OpenAI released GPT-5.6 on July 9 in three variants: Sol at $5/$30 per million tokens (its strongest model), Terra at $2.50/$15 promising GPT-5.5-level performance at half the cost, and Luna at $1/$6 - Engadget. Anthropic answered on July 24 with Claude Opus 5 at $5/$25 per million tokens (unchanged from Opus 4.8), adding an adjustable effort toggle, a fast mode running roughly 2.5x speed at double the base price, and default placement on Claude Max - Anthropic. The launch page's most relevant detail for this guide is a customer evaluation lead reporting Opus 5 scored 9 percentage points higher on their hardest financial-modeling tasks while using a third fewer turns and 60% less time. Every ranked product below that exposes model choice inherits these gains the week it adopts the new engines; every product that hides its engine forces you to guess.

Anthropic's May 5 move deserves its own paragraph because it rewrites what "Claude for spreadsheets" means. The October 2025 finance stack (six Agent Skills, seven connectors) that older articles still describe was replaced by ten prebuilt agent templates: Pitch builder, Meeting preparer, Earnings reviewer, Model builder, Market researcher, Valuation reviewer, General ledger reconciler, Month-end closer, Statement auditor, and KYC screener, alongside new connectors to Dun & Bradstreet, Fiscal AI, Financial Modeling Prep, Guidepoint, IBISWorld, SS&C Intralinks, Third Bridge, and Verisk, on top of the existing FactSet, S&P Capital IQ, MSCI, PitchBook, Morningstar, Chronograph, LSEG, and Daloopa integrations - Anthropic. The same announcement reports Claude Opus 4.7 leading Vals AI's Finance Agent benchmark at 64.37%, superseding the Sonnet-era 55.3% figure that stale rankings still quote. The practical effect: the connector argument that once favored OpenAI's add-in (FactSet and S&P Global as differentiators) is now a wash, because Claude connects to both.

Why does this matter for how you buy? Because distribution and capability have fully converged, and the surviving third-party tools are the ones that are measurably better (GPT for Excel's verified benchmark lead), measurably specialized (Shortcut for modeling, DataSnipper for audit), or measurably cheaper for bulk work (pay-per-use pricing). The generic "AI Excel assistant" wrapping an API for $15 a month is dead as a category, and section 8 walks through the bodies. Your default in August 2026 should be a native agent or a benchmark-proven specialist, and the burden of proof sits on anything that is neither.

2. The Benchmark Reality Check: What the Numbers Actually Say

Ranking these tools used to mean comparing feature lists. That era ended when four serious evaluations landed between late 2025 and mid-2026, giving this category something rare in AI tooling: falsifiable performance data. This section walks through each benchmark, what it measures, who wins, and why the numbers disagree with each other. If you read one section before buying, read this one; a tool's benchmark profile tells you more than its pricing page ever will. For the wider methodology context on how agent benchmarks get designed and gamed, our guide to AI agent evals and benchmarks covers the territory in depth.

The reference benchmark for raw spreadsheet manipulation is SpreadsheetBench, built by the RUC KBReasoning group from 912 questions sourced from real Excel forums, evaluated online-judge style across 2,729 test cases, roughly three variant workbooks per instruction - GitHub. It measures exactly what it says: can the agent produce the precisely correct workbook state for a real user's messy, underspecified request? On the V1-Verified subset (400 real-world tasks, exact-match grading), GPT for Excel scored 92.5% (370 of 400) in a single pass, running Claude Opus 4.7 at medium reasoning, second on the public leaderboard behind a research agent called Tetra-Beta-2 by DealGlass at 94.25% - GPT for Work. Microsoft's own published figure for its Excel agent, from the September 2025 launch announcement, is 57.2%, and the company has not published a newer score since the rename - Microsoft. Google's published figure for Gemini in Sheets is 70.48%, which it describes as approaching human expert capability - Google.

A caution before you treat that chart as one leaderboard: the scores come from different tracks and different dates. GPT for Excel's 92.5% is a May 2026 run on the 400-task V1-Verified subset with a documented setup; Google's and Microsoft's numbers come from their own runs, and Microsoft's predates both the rename and two model-picker upgrades. The honest reading is not "GPT for Excel is exactly 35 points better than Copilot" but that there is a verified, reproducible gulf between the specialist agents and the first-party ones on exact-match manipulation, and the burden is on Microsoft to publish a post-rename score if the gap has closed.

The second data source measures something different: financial modeling judgment. Wall Street Prep asked each tool to build a fully integrated three-statement model for Apple from its latest 10-K using investment banking formatting, then graded the output on accuracy, formatting, and model structure against investment banking standards. The results, in the ranking's June 29, 2026 update: Shortcut 5.9/10, Claude 5.5, Microsoft Copilot 4.4, ChatGPT 2.5, against human analyst benchmarks of 9.4 (top), 7.9 (mid), and 6.4 (low) - Wall Street Prep. Two findings stand out. Even the best AI still scores below the weakest human analyst on a complete modeling task. And speed inverts the picture: the graded tools finished in a fraction of analyst time, which is why banks adopt them for first drafts despite the quality gap.

The third source is about speed at scale, and it produced the most operationally useful numbers in the category. On March 12, 2026, GPT for Work published an 11-task head-to-head against Copilot's agent and Claude in Excel on identical use cases. GPT for Excel won 10 of 11 tests, completing a 10,000-row bulk transformation in 11 minutes 46 seconds; Copilot exceeded 15 minutes without finishing the same job, and Claude in Excel failed to start it under load. On a formula-consistency check, GPT for Excel finished in 16 seconds against nearly five minutes for Copilot, roughly an 18-fold gap, and Claude completed only 60 of 500 rows of a Fortune 500 lookup task before requiring manual termination - GPT for Work. Yes, this benchmark comes from the vendor that won it, so treat the margins skeptically; but the full task list and setup are published, no competitor has issued a rebuttal in the months since, and the rate-limit behavior it documents for Claude matches what users report on long sessions.

The fourth source, AIMultiple's June 16, 2026 benchmark of 15 AI Excel tools, adds breadth: GPT for Excel scored 97.5%, Claude for Excel 95%, and Ajelix 95% on workbook tasks, while Quadratic managed 75% and the long tail collapsed entirely, with ExcelAIBot at 0% and Hugging Face's AI Sheet at 5% - AIMultiple. That floor deserves emphasis: roughly a third of tools marketing themselves for AI Excel analysis fail basic tasks outright, which is exactly why this ranking leans on measured results rather than feature checklists.

It is worth pausing on why exact-match grading is the right standard even though it feels harsh. SpreadsheetBench awards no partial credit for "directionally correct": if the task requires a rule applied to the right range, an off-by-one range fails, exactly as it would fail a real user. Spreadsheets are unusual among AI domains in that outputs are binary at the cell level: a model whose subtotals are 98% correct is not 98% useful, it is broken, because the reader cannot know which 2% to distrust. The multi-workbook judge design also punishes solutions that overfit to one example layout - GitHub. When a vendor quotes a spreadsheet score from a friendlier grading scheme, discount accordingly.

So why do the benchmarks disagree with each other? Because they measure three different skills that correlate less than you would expect. SpreadsheetBench measures edit precision. Wall Street Prep measures domain judgment: does the system understand what a deferred tax liability is and where it belongs. The speed test measures engineering: batching, parallelism, rate-limit handling, which are properties of the product wrapper rather than the model. ChatGPT for Excel is the clearest illustration: OpenAI's own investment-banking benchmark scored its finance-tuned March model at 0.873 against 0.437 for the untuned base model - The Decoder, yet the same product scored 2.5/10 in Wall Street Prep's hands-on grading. A brilliant model in a product that mishandles workbook state still produces a bad workbook. Buy the product, not the model card, and weigh the benchmark that matches your actual workload.

3. The Top 10 AI Agents for Excel Analysis, In Depth

The profiles below expand every row of the master table: what each agent actually is, what the evidence says, what it costs as of August 5, 2026 (verified against live pricing pages the day this refresh was written), and who should pick it. The ordering follows the weighted final scores, which means it deliberately does not follow brand size. Readers of our computer-use benchmark guide will recognize the pattern: measured task completion beats marketing every single time.

One framing note before the list. "Excel analysis" in this guide means the full job: reading a real workbook, understanding its structure, running or building analysis (formulas, pivots, models, transformations), and writing results back without breaking what was there. Tools that only generate formula text, or only analyze an uploaded copy, can still rank if they are excellent at their slice, but the scoring rewards agents that close the whole loop, because that is what the category has become.

#1 GPT for Excel (GPT for Work): The Verified Benchmark Leader

GPT for Excel, built by Talarian under the GPT for Work brand, remains the quiet upset of this ranking: a specialist add-in that outperforms all three frontier labs' native agents on the only exact-match benchmark that matters. Its 92.5% on SpreadsheetBench V1-Verified (370 of 400 tasks, single pass, across lookup, manipulation, and computation categories) is the highest independently documented score for a commercially available Excel agent, second only to the DealGlass research agent at 94.25% - GPT for Work. It backed that up with 97.5% in AIMultiple's June test, the top score among the 15 tools measured - AIMultiple.

The product's real moat is engineering for bulk work. In the March head-to-head it won 10 of 11 tasks and remains the only agent in this list with published evidence of completing 10,000-row transformations without stalling or rate-limiting - GPT for Work. Its record benchmark pass ran Claude Opus 4.7 at medium reasoning, and the product lets you select engines, which matters for the reasons covered in section 5: an engine-transparent product inherits July's model improvements immediately. The trade-off is that it is a third-party add-in from a smaller vendor: no Microsoft-grade compliance story, no bundled identity management, and you will need to clear it with IT in regulated environments.

Pricing - GPT for Work:

The pricing model is the sleeper feature: no subscription at all. Credits stay valid for 12 months after your most recent purchase, each new purchase resets the clock for the whole balance, and credits pool across the team rather than per seat. For an analyst who does heavy spreadsheet work in bursts (month-end, quarter-end), this is dramatically cheaper than a stack of monthly seats that idle most weeks. Best for: analysts and ops teams doing large-volume transformations, enrichment, and cleanup who want the highest verified accuracy per dollar.

#2 Claude for Excel: The Craftsman's Agent, Now With a Real Finance Stack

Claude for Excel is Anthropic's native add-in, and since January 24, 2026 it is included with the standard Claude Pro plan at $20/month, with multi-file drag-and-drop, protection against overwriting existing cells, and automatic context compression for long sessions - The Decoder. What older rankings miss is how much native capability arrived in the February release: pivot table editing (sort, filter, schema changes), chart editing (axes, labels, legends), conditional formatting, sort and filter operations, data validation, and in-sidebar model switching, documented in a hands-on review of the February update - AI Tool Analysis. That review recorded the sidebar offering the then-current Sonnet and Opus generations; Anthropic has not published an updated sidebar model list as of early August 2026, so verify which engines you get on your own tenant before assuming Opus 5 is in there.

The bigger differentiation is the finance stack behind the add-in, rebuilt on May 5, 2026: ten prebuilt agent templates spanning the deal cycle and the close cycle (from Pitch builder and Model builder to General ledger reconciler, Month-end closer, Statement auditor, and KYC screener) plus sixteen-plus data connectors including FactSet, S&P Capital IQ, LSEG, PitchBook, Morningstar, Dun & Bradstreet, IBISWorld, and Third Bridge, with Claude Opus 4.7 leading Vals AI's Finance Agent benchmark at 64.37% - Anthropic. The independent evidence is strong: 95% on AIMultiple's workbook tasks and the second-best score (5.5/10) in Wall Street Prep's analyst-graded ranking, ahead of both Copilot and ChatGPT - Wall Street Prep. Its documented weakness is throughput: in the published speed test it completed only 60 of 500 lookup rows and failed to start the 10,000-row job, so it is the wrong tool for bulk transformations - GPT for Work.

Pricing: $20/month (Claude Pro, includes the Excel add-in), with Max, Team, and Enterprise tiers for higher limits. Anthropic's add-ins now span Word, Excel, PowerPoint, and Outlook with Claude following the user across them - The New Stack. Best for: finance professionals building models, comps, and diligence work who value edit quality and judgment over bulk speed.

#3 ChatGPT for Excel: The Free Default, With EU Fine Print

ChatGPT for Excel is a native add-in that reads and edits the open workbook, launched in beta on March 6, 2026 on a finance-tuned model with FactSet, Moody's, S&P Global, and LSEG data integrations - The Decoder. It reached general availability on May 7, 2026 across ChatGPT plan tiers including Free, alongside a matching Google Sheets add-in - Analytics Vidhya. The "free default" headline now needs two pieces of fine print that almost no ranking carries. First, EU consumer plans are excluded: Plus and Pro subscribers inside the European Union cannot access the add-on, which requires a Business, Enterprise, Edu, or K-12 account there. Second, the free preview for business tiers ended June 2, 2026, after which every add-on interaction consumes plan credits - Pasquale Pillitteri. Free to try remains true for most of the world; free at production volume does not.

The evidence on ChatGPT for Excel stays the most polarized in this list, and understanding why is the key to using it well. On OpenAI's own investment-banking benchmark, the finance-tuned March model scored 0.873 where the untuned base scored 0.437 - The Decoder. Yet Wall Street Prep's hands-on grading scored ChatGPT at 2.5/10, last among the majors, citing structural model errors a first-year analyst would not make - Wall Street Prep. The reconciliation: raw finance reasoning is genuinely excellent, but workbook orchestration (which cell gets which formula, how three linked statements stay linked) was immature at test time. Note also that OpenAI has not published which GPT-5.6 variant, if any, now powers the add-in; the most recent documentation we could verify still records the finance-tuned 5.4-generation engine, so do not assume the July models are inside it yet.

Pricing: Free to try on eligible plans (with plan-level limits and the EU consumer exclusion above), Plus $20/month, with Business, Enterprise, and Edu tiers carrying credit-metered add-on usage. Best for: everyday analysis, data questions, and one-off transformations at zero marginal cost, with human review on anything structural.

#4 Gemini in Google Sheets: The Other Ecosystem's Champion

Gemini in Google Sheets is not an Excel tool, and it earns its place in this ranking anyway, for two reasons. First, an enormous share of "Excel analysis" work actually lives in Sheets, especially at startups and in ops teams. Second, Google published the strongest first-party benchmark score in the category: 70.48% on SpreadsheetBench, announced March 10, 2026, which it describes as approaching human expert capability - Google. That beats Microsoft's published figure by more than 13 points on the same benchmark family.

The agentic capability arrived in force on April 22, 2026: Gemini can build entire spreadsheets from a single prompt, including data retrieval and formatting, and refine existing models through side-by-side editing, with Workspace Intelligence synthesizing data across files, emails, and chats - Google Workspace Updates. Two rollout details matter for buyers this month. The launch ran as a promotional access period through July 15, 2026, initially US-only in English, after which access follows the standard edition entitlements (Business Standard and Plus, Enterprise Standard and Plus, and the Google AI consumer plans). And Google keeps shipping internationalization fast: Fill with Gemini gained 11 additional languages on July 7, 2026, bringing structured AI fill to nineteen languages total - Google Workspace Updates. The limitations are the mirror image of the strengths: it lives in Google's ecosystem, .xlsx files round-trip through import, and complex Excel-native features (macros, some pivot configurations) do not translate.

Pricing: included with Google Workspace editions that carry Gemini and with Google AI consumer plans; for most Workspace organizations there is no incremental per-seat cost, which makes its value score hard to beat. Best for: teams already on Workspace, and anyone whose "spreadsheet analysis" is really Sheets-based ops, marketing, or reporting work.

#5 Shortcut: The Financial Modeling Specialist

Shortcut (shortcut.ai), built by Fundamental Research Labs, is what happens when a team builds an Excel agent for exactly one demanding audience: finance professionals. It is the #1 ranked tool in Wall Street Prep's evaluation at 5.9/10, ahead of Claude (5.5), Copilot (4.4), and ChatGPT (2.5), on a full three-statement Apple model built from 10-K filings to investment banking standards - Wall Street Prep. Even at #1, Wall Street Prep's graders note it underperformed a lower-bucket human analyst, which is the honest ceiling of the whole category right now: the best AI modeler money can buy produces a strong first draft, not a finished deliverable.

Since our last refresh the product widened its footprint in two verifiable ways: the free tier is now 20 free weekly credits (not the one-time trial older coverage describes), and the platform spans web, desktop, an Excel plugin, and a Google Sheets plugin - Shortcut. Handle the vendor's own bolder marketing claims (beating first-year bankers head-to-head in self-run trials) with care: the methodology is the vendor's own, and the independently graded result is the more trustworthy signal. It is impressive enough on its own.

Pricing - Shortcut:

Messages typically cost 2-15 credits each, so the free tier genuinely supports light weekly use and the Pro plan supports a serious monthly workload. At 5x the price of Claude Pro, it only makes sense if financial modeling is your actual job. Best for: investment banking, private equity, and hedge fund analysts who want the strongest graded model-building agent and can justify $100/month against billable time.

#6 Edit with Copilot (formerly Agent Mode): The Incumbent Under Pressure

Edit with Copilot is the agent most people will encounter first, because it ships inside Excel itself. This is the feature Microsoft launched as Agent Mode in September 2025, took to desktop general availability on January 27, 2026, and renamed in March 2026, when it also gained the ability to work on local files with no OneDrive or SharePoint requirement - Chris Menard Training. Its strategically boldest feature survives the rename: a model switcher letting users run workbook tasks on OpenAI or Anthropic models inside a Microsoft product, with an Auto mode that picks for you - Microsoft Tech Community. Integration is its unassailable advantage: nothing to install, tenant-level security and compliance, and IT departments already know how to govern it. One regional caveat from the same GA post: consumer Personal and Family subscriptions run on an AI credit model, and the agent was not yet available to EU and UK consumers at desktop GA.

The measured performance is where the story sours. Microsoft's own published SpreadsheetBench figure is 57.2%, disclosed at the September 2025 launch and never updated since - Microsoft. Wall Street Prep graded it 4.4/10, below Shortcut and Claude. And in the published speed head-to-head it exceeded 15 minutes without finishing the 10,000-row job that GPT for Excel completed in under twelve, with an 18-fold gap on a formula-consistency task - GPT for Work. None of this makes it useless; it is genuinely good at conversational analysis, summaries, and moderate transformations inside a governed enterprise environment, and the local-files change plus the dual-lab picker mean its ceiling rises whenever either lab ships a better engine. But the data says it is currently the weakest of the big-three native agents at exactly the heavy analytical work this guide ranks, and its adoption case rests on distribution and governance rather than capability.

Pricing - Microsoft:

Best for: enterprises that need governed, compliant AI inside the Office estate and can accept mid-pack analytical performance; for how Microsoft's broader agent push fits together, see our Copilot Cowork analysis.

#7 DataSnipper Excel Agents: The Audit Specialist

DataSnipper has spent years as the default Excel-embedded toolkit in audit, and its purpose-built Excel Agents for audit and finance offer the one property generic agents cannot: evidence traceability. Every result can be traced back to the originating document and location, supporting the transparent review and documentation that assurance work requires; the agents execute structured procedures directly in workbooks rather than offering advisory suggestions - DataSnipper. DataSnipper's own framing of the contrast is instructive: generic tools produce suggestions that still need manual documentation, while its agents produce audit-ready workpapers with governance built in.

This entry illustrates a structural truth about the 2026 agent market: once frontier capability becomes a commodity input, defensibility moves to workflow specificity. A Big Four audit team cannot submit workpapers where an AI's number has no documented provenance, no matter how high the underlying model's benchmark score. DataSnipper accepts a lower ceiling on general analytical brilliance in exchange for outputs that survive professional review, regulatory inspection, and litigation discovery. The same pattern shows up across regulated industries, as we documented in how the financial sector automates with AI agents. Because no public benchmark covers assurance-grade agents, its accuracy score in the master table is a judgment call based on workflow fit, and we label it as such rather than dressing it up as measurement.

Pricing: enterprise quotes only, seat-based, typically procured firm-wide. That opacity costs it points in the value column, but audit buyers do not shop on price pages. Best for: audit, controllership, and SOX/compliance teams where every number must be traceable to evidence.

#8 Ajelix: The Workbook Power Toolkit, Repriced

Ajelix built its reputation as a budget toolkit (formula generation, VBA scripting, workbook analysis), and its accuracy has genuinely kept pace with the frontier: it scored 95% on AIMultiple's workbook tasks, matching Claude for Excel on that test - AIMultiple. Its VBA and script generation remain among the best anywhere, and for teams maintaining legacy macro-heavy workbooks it solves problems the newer native agents still ignore.

The catch is the pricing regime it moved to. The old $15-20 toolkit plans are gone, replaced by per-seat agentic tiers: Lite at $39/user/month ($32 billed yearly), Pro at $99/user/month ($82 yearly), and Max at $199/user/month ($165 yearly) - Ajelix. That lands Ajelix in an awkward position: at $39-99 per seat it competes with Claude Pro at $20 and Copilot Business at $18, both offering broader capability, while its accuracy edge over them on workbook tasks is a rounding error. The value column reflects that squeeze. Best for: teams with heavy VBA and macro estates and specific workbook-automation needs that the native agents handle poorly, who can justify the per-seat cost against saved engineering time.

#9 Julius AI: The Conversational Data Analyst

Julius AI represents the upload-and-chat generation of tools, and it remains the best of that breed for statistical work. You drop in files (Excel, CSV, and more), and Julius writes and executes real analysis code, producing charts, regressions, and forecasts through conversation. It behaves like a junior data scientist rather than a spreadsheet macro: it runs the ANOVA, explains its assumptions, and shows the code, which makes its work inspectable in a way formula edits are not. That audit trail is why researchers keep it alongside a native agent rather than instead of one.

Two honesty flags earn it a lower row this year. First, it is absent from the 2026 Excel benchmarks this guide leans on, so its accuracy score rests on its established notebook-analysis reputation, not on measured workbook performance. Second, its pricing is genuinely hard to verify: julius.ai blocks automated access, so the figures below come from a third-party pricing guide dated April 20, 2026 and should be treated as reported rather than confirmed. That guide lists Free, Plus at $35/month, Pro at $45/month (unlimited messages, database connectors), Max at $200/month, and Business at $375/month, with roughly 15-17% off annually and a 50% student discount - Coefficient. Structurally, Julius analyzes copies of your data, not your live workbook: no in-cell editing, no formula preservation, no round-trip fidelity for a complex model. For pure analysis that is irrelevant; for workbook repair it is disqualifying. Best for: students, researchers, and analysts who want conversational statistics and visualization on exported data, with methodology on the record.

#10 Formula Bot: The Repositioned Veteran

Formula Bot (formulabot.com) was the archetypal narrow tool of the last generation: type what you want, get a formula. The product has since repositioned as a full "AI data analyst" with data connectors, chat-with-your-data, automated dashboards, and playbooks, at Starter $18/month (250 chats), Max $29/user/month (unlimited chats), and Enterprise $149/user/month, with roughly 30% off yearly - Formula Bot.

It earns the last ranked spot on honest grounds: there is no independent benchmark evidence for its analytical accuracy (it appears in no major 2026 test we could find), and its analysis happens outside your live workbook. But the formula-and-explanation core remains genuinely useful, the price is among the lowest in the category, and for a non-technical user who mostly needs plain-English-to-formula plus occasional CSV chat, it delivers exactly that without the cognitive overhead of a full agent. Best for: casual and non-technical spreadsheet users with formula-centric needs and light analysis, at minimal cost.

Honorable Mentions and the Entrants We Refuse to Rank

Two tools narrowly missed the table on evidence grounds rather than quality grounds. Quadratic is a spreadsheet rebuilt from scratch around code and AI (Python, SQL, and formulas in one grid); it scored 75% on AIMultiple's workbook benchmark, respectable for a product that is not Excel at all, and it points at where technical analysts may land when they stop pretending Excel is an IDE - AIMultiple. Numerous.ai occupies the bulk-functions niche at about $10/month billed annually (with a $1 seven-day trial) for roughly 500,000 characters of AI input/output and 500 formula generations across Excel and Google Sheets - Numerous. It is the cheapest way to run "categorize these 2,000 rows" tasks if you do not need a full agent.

Then there is the category we deliberately exclude, and naming it is part of this guide's job. The 2026 search results for this exact query now include multiple freshly launched compare pages whose own product sits at #1 on the strength of a self-run accuracy figure with no published methodology, no task list, and no third-party replication. We do not rank those entrants, not because they are necessarily bad products, but because an unverifiable self-benchmark is not evidence, and a ranking that launders it into a scoreboard becomes part of the problem. The inclusion bar for this list is simple and worth stealing for your own procurement: a named independent measurement, a live verifiable price, or both. Tools clearing neither bar compete for your attention with SEO, and SEO is not a spreadsheet skill.

4. Head-to-Head: Edit with Copilot vs Claude vs ChatGPT vs Gemini

Most buying decisions in 2026 come down to the four native agents, so they deserve a direct comparison on identical axes. The table compresses what the preceding sections established, and the prose after it covers the second-order differences that do not fit in cells. If your organization standardizes on one of these, this is the section to forward.

The most consequential row is model choice, because it flips the usual platform logic. Microsoft, of all vendors, is the only one offering a cross-lab picker: the same workbook task can run on an OpenAI or an Anthropic model, with an Auto mode choosing on your behalf - Microsoft Tech Community. That is a hedge against exactly the divergence the benchmarks show (OpenAI models leading on finance reasoning tests, Anthropic models leading on hands-on modeling quality), and it means Copilot's ceiling rises automatically whenever either lab ships a better engine, including July's releases. Anthropic's counter is depth over breadth: its add-ins follow the user across Outlook, Word, Excel, and PowerPoint, so a diligence session can flow from a data-room document to a model to a memo without re-explaining anything - The New Stack.

The connector row changed materially since spring, and it changed in Anthropic's favor. Where early 2026 coverage framed FactSet and S&P Global as OpenAI differentiators, Claude's May 5 stack now includes both, plus a private-markets and credit lineup (PitchBook, Chronograph, Dun & Bradstreet, Verisk, IBISWorld) that reads like a buy-side analyst's vendor list - Anthropic. Microsoft's answer remains structurally different: Copilot's data advantage is not market data but your organization's own data via the Graph (meetings, mail, files), which is a complementary kind of context. If your analysis depends on licensed financial data, check the connector list before the benchmark score; an agent that cannot reach your data sources scores zero on your actual workflow.

The second-order difference is who each vendor thinks the user is. OpenAI's free-tier GA is a consumer-scale land grab, now with visible monetization seams (EU consumer exclusion, credit metering since June 2). Anthropic is running a professional-tools play: agent templates, connectors, edit quality, at a professional's price. Microsoft is selling to the CIO: governance, tenant compliance, bundling. Google is selling to the organization that already left Excel. None of these strategies is wrong, and the practical takeaway is that your procurement context (who pays, who governs, what ecosystem you live in) legitimately changes which agent is right, independent of raw scores.

5. Models Under the Hood: Opus 5, GPT-5.6, and Why the Engine Matters

A year ago, asking which model powers an Excel tool was trivia. In August 2026 it is a first-order buying question, for a structural reason: the products have converged on similar interfaces (a chat pane plus direct workbook access), so the remaining variance concentrates in the model and the orchestration around it. Two products with identical feature lists can differ by 30 benchmark points because one runs a frontier reasoning engine with careful workbook-state management and the other runs a cheap model with none.

The July generation change is the news. Claude Opus 5 shipped July 24 at $5/$25 per million tokens, the same price as Opus 4.8, with an adjustable effort setting to trade intelligence against token spend and a fast mode at roughly 2.5x speed for double the base price; it is the default on Claude Max and the strongest model on Claude Pro, and the launch page carries a customer report of 9 points higher accuracy on hard financial-modeling tasks with a third fewer turns - Anthropic. GPT-5.6 shipped July 9 in three variants: Sol ($5/$30) as the flagship, Terra ($2.50/$15) matching GPT-5.5 performance at half the cost, and Luna ($1/$6) for volume work - Engadget. We break the head-to-head down in detail in our GPT-5.6 vs Claude Opus 5 comparison, with deeper single-model coverage in the Opus 5 vs 4.8 benchmark guide and the GPT-5.6 benchmark and pricing breakdown.

Here is where each ranked product actually stood on engines at the time of writing, and this is a place where honesty beats confidence. GPT for Excel ran its record SpreadsheetBench pass on Claude Opus 4.7 at medium reasoning and exposes engine choice to users, so July's models are a settings change away - GPT for Work. Edit with Copilot's picker offered the then-latest OpenAI and Anthropic models at its January GA, with Microsoft positioning the Auto mode to route on your behalf - Microsoft Tech Community. Claude for Excel documented in-sidebar switching between its Sonnet and Opus generations in the February release; whether the sidebar now offers Opus 5 is not publicly documented as of August 5, so check your tenant - AI Tool Analysis. ChatGPT for Excel's most recently documented engine is the finance-tuned 5.4 generation; OpenAI has not published a GPT-5.6 upgrade for the add-in, so we do not claim one - Pasquale Pillitteri. Vendors that go quiet about engines after a model cycle are usually running the old one.

Why should a spreadsheet user care about token prices and effort toggles? Because they translate directly into workbook-scale behavior. Rate limits and per-token economics explain the observed failure modes from section 2: Claude's 60-of-500-rows stall is a throughput economics problem, not an intelligence problem, and GPT for Excel's bulk dominance is largely batching engineering on top of the same class of models. Opus 5's effort toggle and Terra's half-price positioning both push in the same direction: cheaper mid-tier reasoning for volume tasks, which is exactly what bulk spreadsheet work consumes. Teams routing different workloads to different engines are already exploiting this, a pattern we quantify in our model routing cost guide. The practical buying rule is simple: prefer products that disclose and let you choose the model, because those are also the products that improve the week a better engine ships. For a broader view of which engines currently lead for agentic work overall, see our August 2026 LLM-for-agents ranking.

One dated but instructive data point survives from the March cycle: OpenAI's own investment-banking benchmark scored its finance-tuned model at 0.873 against 0.437 for the same-generation untuned base - The Decoder. Domain tuning roughly doubled finance-task performance within one family. That is the same lesson Wall Street Prep teaches from the product side, and it explains why "which engine, tuned how, orchestrated how" is now inseparable from tool choice.

6. Accuracy, Failure Modes, and Human Oversight

The benchmark numbers in this guide range from 57% to 94%, and the single most important skill for working with these agents is knowing what those percentages mean operationally. A 92.5% exact-match score means that on roughly one task in thirteen, the agent's output was wrong in a way exact-match grading catches. A 57% score means wrong outcomes on two tasks in five. Neither number is a reliability guarantee, and both were achieved on benchmark tasks with defined correct answers; your ambiguous, path-dependent, legacy-workbook reality is harder.

The failure modes cluster into three families, each with a different mitigation. Silent numerical errors are the most dangerous: a formula referencing the wrong range, a hardcoded value where a link should be, a sign flip in a cash flow. Exact-match benchmarks catch these; your eyes might not, because the output looks plausible. Wall Street Prep's finding that every AI scored below its lowest human analyst benchmark (best AI 5.9 against a 6.4-9.4 human range) is fundamentally a statement about this failure family - Wall Street Prep. Scale and rate-limit failures are the second family: Claude completing 60 of 500 rows, Copilot's agent never finishing the 10,000-row job. These at least fail loudly, but they fail after consuming your time, so match the tool to the workload size before starting - GPT for Work. Destructive edits are the third: overwriting formulas with values, breaking links between sheets. This is why Anthropic shipped explicit overwrite protection and why version history should be non-negotiable in your workflow.

From those families, a practical oversight protocol follows. Keep it lightweight or nobody will follow it:

• Version everything: the agent works on a copy or with tracked changes, never the only copy
• Tie out totals: independently verify 3-5 anchor numbers against source data after any agent pass
• Audit formulas, not values: scan for hardcodes where formulas should be
• Match tool to scale: bulk jobs go to bulk-proven tools, judgment jobs to judgment-proven tools
• Escalate by stakes: anything that leaves the building gets full human review

A concrete composite example makes the stakes tangible. An FP&A analyst asks an agent to update the revenue model with Q2 actuals and refresh the full-year forecast. The agent pastes the actuals correctly, updates eleven of twelve driver formulas, and on the twelfth quietly replaces a growth-rate formula with last quarter's computed value: a hardcode. Every total still calculates, the sheet looks refreshed, and the forecast is now insensitive to the very assumption it is supposed to flex on. Nothing errors. The bad number surfaces three weeks later in a board scenario discussion when someone asks why the downside case barely moves. This is the archetypal agent failure: plausible, silent, and downstream. The two habits that catch it are the formula audit (a thirty-second hardcode scan with trace precedents) and anchoring a handful of totals against the source export. Teams that institutionalize those two habits capture nearly all of the speed win at a small fraction of the risk.

The deeper principle is that these agents have crossed the threshold where they are faster than you but not yet more reliable than you. A model draft produced in a fraction of analyst time is real value even after you spend twenty minutes reviewing it. The trap is letting the review shrink because outputs usually look right: "usually" at 92.5% means a bad number ships every couple of weeks of daily use, and at 57% it means one ships most days. Calibrate review effort to the tool's measured accuracy and the task's blast radius, not to your impression of the last few outputs, a discipline whose economy-wide version we explore in our true cost of agentic AI report.

7. Which Agent for Which Job: A Persona-Based Selection Framework

Rankings answer "which is best overall," but nobody buys overall. You buy for a workload, and the evidence in this guide maps cleanly onto four workload archetypes. Find yours below, and treat the ranking's top spot as a tiebreaker rather than a mandate. The decision logic compresses into one diagram.

The finance analyst (banking, PE, corporate FP&A) should start with Claude for Excel at $20/month: second-best analyst-graded quality, ten prebuilt finance agents, and a connector stack that now covers FactSet, S&P Capital IQ, and the private-markets vendors out of the box - Anthropic. If model building is the core of your role and $100/month clears against billable hours, Shortcut is the graded quality leader, and its 20 free weekly credits make trying it on one live deal essentially free. Run both before committing; their failure styles differ (Claude occasionally under-builds, Shortcut occasionally over-structures), and which one annoys you less is personal.

The auditor or controller has a shorter list, because the constraint is not intelligence but provenance. DataSnipper's Excel Agents are the only ranked option whose outputs natively carry evidence links suitable for workpapers - DataSnipper. Use a general agent for internal analysis if you like, but keep it out of anything that faces a regulator or an audit partner until traceability catches up across the category.

The ops, marketing, or data team doing bulk work should optimize for throughput per dollar, where the published evidence is unambiguous: GPT for Excel is the only agent with documented 10,000-row completion, and its pay-per-use credits mean idle months cost nothing - GPT for Work. If your stack is Google-side, Gemini in Sheets turns bulk classification and enrichment into structured fill operations, now in nineteen languages - Google Workspace Updates. Numerous.ai at about $10/month remains the budget pick for light recurring bulk tasks.

A note on piloting method before the last persona, because most evaluations in this category are decided by demos, and demos systematically mislead. Vendors demo on clean data with well-specified requests; your reality is merged cells, inconsistent date formats, and instructions like "make this match how Sarah did it last quarter." The fix is cheap: assemble a five-task pilot set from your team's actual recent work (one bulk transformation, one model edit, one analysis question, one formatting task, one deliberately ambiguous request), run each candidate on the identical set, and grade exact-match style: did it produce the workbook state you actually needed, and how long did review take? The free tiers now make this nearly costless: ChatGPT for Excel outside the EU consumer carve-out, Shortcut's 20 weekly credits, Claude bundled in a $20 Pro seat, Copilot's consumer trial path at $19.99. Two hours of this beats any ranking, including this one, because it measures the only benchmark that matters: your workload.

The everyday business user should simply start free where eligible: ChatGPT for Excel costs nothing to evaluate against your real spreadsheets on most plans outside the EU consumer tiers - Analytics Vidhya, and Microsoft 365 Premium at $19.99/month or Copilot Business at $18/user is the governed path if your organization lives in Microsoft's stack - Microsoft. The dashed arrow in the diagram matters more than any single pick, though: if your "Excel task" is really fetch data, then analyze, then email the result, no in-cell agent covers that loop, and section 9 is about what does.

8. Deadpool and Rename Watch: Rows, Akkio, Copilot Pro, Agent Mode

Any ranking published in early 2026 recommended tools that are now acquired, repositioned, retired, or renamed. This section exists because stale recommendations are expensive: adopting a platform being folded into another product means migrations, retraining, and re-procurement within a year, and searching for a retired feature name means missing current documentation entirely. Here is what changed, and what each case teaches about the category's structure.

The biggest exit: Rows was acquired by Superhuman, announced February 23, 2026, with the team and its AI-analyst technology folded into Superhuman Docs (the product formerly known as Coda) - Superhuman. The rows.com pricing page still displays legacy Plus ($8/user/month, $6 annual) and Pro ($79/month + $8/user) tiers under a "Rows joined Superhuman" banner - Rows, but recommending platform adoption of Rows in August 2026 is recommending a product in wind-down. If you are on it, plan your migration on your own schedule rather than Superhuman's.

Three quieter changes complete the picture. Akkio, which earlier lists ranked as a general predictive-analytics pick, now positions itself entirely as "the AI platform that automates campaign workflows" for media agencies, with clients like Havas, Dentsu, and Horizon Media and no public per-seat pricing - Akkio. It is not dead; it simply no longer belongs on a general Excel-analysis list. Copilot Pro, the $20 consumer plan many 2025 guides recommended, is gone from Microsoft's consumer lineup, its role absorbed by Microsoft 365 Premium at $19.99/month - Microsoft. And the Agent Mode name itself joined the deadpool in March: same capability, new "Edit with Copilot" branding, plus local-file support - Chris Menard Training. A reader can reasonably conclude that in this category, names rot as fast as prices.

There is also a quieter churn that makes no acquisition headlines: pricing regime changes that transform a tool's value proposition without changing its logo. Ajelix's move from budget toolkit plans to $39-199 per-seat agentic tiers, Formula Bot's repositioning into an $18-149 data-analyst ladder, Julius's reported climb to a $35-45 mid-tier, and OpenAI's June 2 shift from free preview to credit-metered add-on usage all happened within roughly two quarters. The pattern is rational: as the frontier labs commoditize the low end (a free formula-generating tier destroys the $15 formula-generator market overnight), independent vendors either move upmarket with agentic claims and per-seat pricing or retreat into a niche the labs will not chase. For buyers, the operational consequence is that any price you saw more than a quarter ago is a rumor. Re-verify at renewal, and treat a sudden upmarket repricing as a signal to re-run your evaluation, because it usually means the vendor's old segment stopped paying.

The structural lesson ties back to section 3's exclusions. The middle of this market cannot hold: below the frontier labs and above the cheap niche utilities, a standalone AI spreadsheet platform needs a proprietary distribution wedge, a verified performance lead, or a workflow requirement to justify existing. Rows had none at sufficient scale and sold. Akkio retreated to a vertical where it has a wedge. The newest entrants skip evidence entirely and compete on programmatic SEO with self-scored benchmarks, which is the same structural squeeze expressing itself as content instead of product. The survivors on this list own the surface (Microsoft, Google), the model (OpenAI, Anthropic), a verified lead (GPT for Work), or a workflow requirement (DataSnipper, Shortcut). Evaluate every new entrant against that test.

9. Beyond the Spreadsheet: When Excel Analysis Is Just One Step

Here is the question none of the ten ranked tools can answer: what happens before and after the spreadsheet? Real analytical work rarely begins with a clean workbook and ends with a computed cell. It begins with "pull this quarter's numbers from three portals and our CRM," and it ends with "send the variance summary to the leadership list and file the workbook in the deal folder." Every agent in this ranking operates inside the grid; the workflow around the grid is left to you, and in most teams that surrounding work takes longer than the analysis itself.

We can speak to this one from direct operational experience rather than benchmark reading. At O-mega, our agent workforce's computer sessions generate, transform, and analyze real .xlsx files daily as one step inside longer autonomous runs, and the failure pattern we observe matches the benchmarks in a specific way: agents rarely fail at the arithmetic, they fail at workbook state and workflow glue, at knowing which file is current, what upstream export changed, and where the result needs to land. That is why the in-cell agents above and workflow-level platforms are complements, not substitutes. An O-mega agent can log into a supplier portal, download the export, run the transformation, and distribute the result by email, treating the spreadsheet step as one link in a chain; the mechanics of agents driving real web interfaces are covered in our browser-use agent review, and the surrounding desktop-and-computer layer in our agentic computer use guide.

The economics of the two categories differ in a way that matters for budgeting. In-cell agents price per seat or per token because they amplify a human who stays in the loop for every session; workflow platforms price closer to outcomes because the human leaves the loop for entire runs. That difference compounds on recurring work. A weekly competitor pricing report that takes an analyst two hours (gather, paste, analyze, summarize, send) costs roughly a hundred analyst-hours a year even when the analysis step is agent-accelerated to minutes, because the gathering and distribution steps still consume the human. Automating the full loop converts that recurring cost into review-only time: the analyst reads the finished report and spot-checks anchors, which is section 6's oversight discipline applied one level up.

The first-principles way to decide between an in-cell agent and a workflow agent is to count the surfaces your task touches. One surface (the workbook) means an in-cell agent from this ranking is the right tool: it has the deepest possible integration with that surface. Three or more surfaces (a portal, a workbook, an inbox) means the bottleneck is orchestration, not spreadsheet intelligence, and an in-cell agent leaves you as the human glue between steps. Recurring multi-surface work is exactly the shape that workforce-style platforms automate end to end. Many teams sensibly run both: Claude or GPT for Excel for deep interactive workbook sessions, and an autonomous platform for the recurring pipelines those workbooks feed.

10. Conclusion: The Decision Framework

This category now updates faster than its own literature. Since our January version: Microsoft renamed its Excel agent and unlocked local files, Anthropic rebuilt its finance stack around ten agent templates and sixteen-plus connectors, OpenAI took its add-in GA and then started metering it, Rows sold, Copilot Pro died, and both frontier labs shipped a new model generation inside fifteen July days. The refresh you are reading replaced or re-verified essentially every number in the previous version, which is the clearest signal of the market's velocity we can offer.

The decision framework distilled: default to a native agent and pay for a specialist only when the evidence says your workload demands it. Start with ChatGPT for Excel free where eligible to establish a baseline. Upgrade to Claude for Excel ($20) for finance-grade edit quality and the agent-template stack, to GPT for Excel (pay-per-use) when bulk scale or maximum verified accuracy is the job, or to Shortcut ($100) when graded modeling quality is worth five times the price. Choose Edit with Copilot ($18-32/user, $19.99 consumer) when governance and Microsoft-stack integration outweigh mid-pack benchmark performance, Gemini in Sheets when your organization lives in Workspace, and DataSnipper when every number needs a paper trail. And keep every agent's measured accuracy taped above your monitor: at 57-92%, these tools are extraordinary accelerators and unacceptable final authorities.

Watch three things for the rest of 2026. Whether Microsoft publishes a post-rename SpreadsheetBench score, because a dual-lab model picker riding Opus 5 and GPT-5.6 should close some of the 35-point gap, and silence will say as much as a number. Whether the July engines pull exact-match scores above 95% and start collapsing the review burden that section 6 describes. And whether the boundary between in-cell agents and autonomous workflow platforms keeps blurring, because the vendor that credibly does both (deep workbook intelligence plus end-to-end orchestration) will make every ranking in this guide obsolete again.

This guide was written by Yuma Heymans (@yumahey), founder and CEO of O-mega and co-founder of HeroHunt.ai, whose agents produce more spreadsheets in a week than he has patience to open, which is precisely why the verification discipline in section 6 exists.

This guide reflects the AI Excel agent landscape as of August 5, 2026. Every price and benchmark was checked against a live source page on that date. Pricing, model versions, and leaderboards in this category change monthly: verify current details on vendors' live pages before purchasing.