Best AI for business

"Business work" covers drafting reports and slides, working through spreadsheets, answering customer requests and running multi-step workflows across tools such as a CRM, a help desk and a finance system.

Who this page is for

This page is meant for teams choosing a model for business work, whether in a chat app or through an API: knowledge work that ends in a document, slide deck or spreadsheet, and automated workflows that move data between business apps.

In the app or via API? This ranking scores the models behind both, but its price criterion is the API price per token. If your team works in a chat app instead, see the AI subscriptions scored for work, which compares consumer plans per price tier on model quality, usage, price and work features.

How this ranking works

The ranking combines four criteria into one score. The weights are fixed and add up to 100%. Our methodology explains how each criterion is scaled and combined.

• GDPval-AA v2.1, 40%. Artificial Analysis rates models on 220 knowledge-work tasks across 44 occupations that need documents, slides, diagrams or spreadsheets as output. Models work in an agentic loop with shell access and web browsing, and an LLM judge compares outputs in pairs to produce Elo ratings ^[1]. It is the closest match to everyday office deliverables, so it gets the largest weight. Higher is better.
• AutomationBench v1.0.6, 25%. Zapier tests agents on end-to-end workflows in sales, marketing, operations, support, finance and HR, using simulated business apps. Scoring is deterministic and checks only the final state, with no LLM judge ^{[2] [3]}. Higher is better.
• Blended price, 25%. Business workloads often run at volume, so the list price per token matters. Lower is better, on a log scale, so that a few very expensive models don't flatten the differences among the rest.
• Tool calling, 10%. Business automation depends on calling tools and APIs. Most models in our catalog support it.

Which benchmark results count

Both benchmarks list the same model several times, for example at different reasoning efforts or with a fallback model. For this ranking each benchmark criterion uses a setting group, a list of stored settings we treat as comparable:

• GDPval-AA: all v2.1 runs count, at any reasoning effort and including non-reasoning runs. This covers the Stirrup agent harness that Artificial Analysis uses, and the Anthropic runs with "Default Fallback", where refused steps go to a fallback model ^[1].
• AutomationBench: all v1.0.6 leaderboard runs count, at any reasoning effort, including runs where refused tasks go to a fallback model ^[2].

Within each group, a model counts with its best run. Independent results are preferred over results a provider reports itself; if only a provider result exists, it is used and marked "self-reported". The methodology box on this page lists every stored setting in each group.

A model needs data for at least 60% of the total weight to be ranked. Price and tool calling together make up 35%, so a model needs a result on at least one of the two benchmarks.

OSWorld 2.0, which tests computer-use agents on long, multi-step desktop workflows ^{[4] [5]}, is not part of the ranking. Its results are only comparable when the release, task set, metric, step budget and agent approach all match, and most current numbers are reported by the providers themselves ^[4]. You can still see them below.

How to read the results

• Scores are relative. Each criterion is rescaled from 0 to 100 across the candidates on this page. 100 means best among these models, not perfect, and adding or removing a model can shift the others.
• Check the setting. The table shows the setting behind each score, such as the reasoning effort or a fallback run.
• "Self-reported" marks a result that comes from the model's own provider rather than from Artificial Analysis or Zapier.
• Small gaps are noise. Elo ratings such as GDPval-AA's move when models are added ^[1]. Treat models close in score as roughly equal.

Limitations

• Grouped runs differ in reasoning effort, and some include a fallback model. A fallback run partly reflects the fallback model, not only the model being ranked.
• Because each model counts with its best run, a model tested at many effort levels has more chances to score well than one tested once.
• Benchmark numbers change. Leaderboard versions are not comparable with each other, so we only use GDPval-AA v2.1 and AutomationBench v1.0.6 ^{[1] [2]}.
• Some results in our data are reported by the providers themselves; the tables label them "Self-reported".
• Only models with a result in at least one of the two groups can be ranked. A missing result doesn't mean a model is weak, only that we have no comparable measurement.
• None of these benchmarks covers data protection, contract terms or where your data is processed, which often decide business purchases.
• The data behind this page is as of October 2026.

All published results

The benchmark pages list every result we hold, with its setting and who reported it: GDPval-AA v2.1, AutomationBench and OSWorld 2.0.

How to choose for your own work

The ranking narrows the field; your own requirements decide.

• Look at the criteria, not only the total. A model strong on GDPval-AA suits document and analysis work; a model strong on AutomationBench suits workflows across apps. The table shows both.
• Check prices. Business workloads often run at high volume, so cost matters. Use AI pricing and the cost calculator with your own token counts.
• Compare a shortlist. Put two to four candidates side by side in compare and check context window, tool support and price together.
• Test on your own tasks. A few dozen real requests from your team say more about fit than any single benchmark. How to compare AI models explains how to set up a fair test.

Related guides

• How to choose an AI model
• How to compare AI models
• AI token pricing explained
• How to reduce AI API costs
• Reasoning models explained