Best AI for Business Work in 2026
One score for knowledge work, built from four benchmarks: GDPval-AA, AA-Briefcase, the Vals Finance Index and the Vals Legal Index. It covers 36 models, including the ones that only some benchmarks have tested.
Last updated: · 36 ranked
The short answer
Updated October 2026Claude Opus 5.5 (Anthropic) and Muse Wonder Full (Meta) are effectively tied at the top, with Claude Sonnet 5.5 next. The top two are separated by a gap of 1.0 points, inside the margin of error, so treat them as tied. For the best result per dollar, DeepSeek V4.1 Flash 1017 (DeepSeek) scores 62 at $0.26 per task, about 23× cheaper than Claude Opus 5.5. Full ranking below.
- 1 Claude Opus 5.5 Work Score 72
- 2 Muse Wonder Full Work Score 71
- 3 Claude Sonnet 5.5 Work Score 70
How the EVY Work Score works
Each benchmark tests a different slice of professional work, and none of them has tested every model. Averaging the scores a model happens to have would reward models that skipped the hard tests. The Work Score fixes that in three steps:
- One scale. Elo ratings and percentages are converted to the same units, so they can sit side by side.
- Difficulty adjustment. The calculation learns how hard each benchmark is and how well it separates strong models from weak ones. Doing well on a hard benchmark counts for more.
- Gaps filled by estimate. A model's score is the ability level that best explains the results it has. Models with fewer results are pulled toward the average and shown with a wider margin, so one lucky number cannot put a model on top.
The benchmarks agree with each other: on models that have both, their scores correlate between 0.78 and 0.96. That agreement is why one combined number is meaningful. A score of about 50 is average across the models tracked here, and the ± after each score is one standard error. Gaps smaller than that margin are ties.
Show benchmark details Hide benchmark details
- GDPval-AA Artificial Analysis runs OpenAI's GDPval tasks: real work from 44 occupations in 9 industries. Two anonymised answers are compared by an LLM judge and aggregated into an Elo rating. 27 models. About 96% of its variation is explained by general work ability.
- AA-Briefcase Artificial Analysis runs a private test of long, multi-step knowledge work: realistic business workflows that end in deliverables such as spreadsheets, presentations and memos. Each model is scored at its best effort setting. 33 models, matched to the models tracked here. About 94% explained by general work ability.
- Vals Finance Vals AI combines Finance Agent, EMB and Tax Agent tests into one accuracy score. 30 models. About 87% explained by general work ability.
- Vals Legal Vals AI combines case-law questions across practice areas with legal research and agent-style legal tasks. 27 models. About 75% explained by general work ability.
Known limits: GDPval-AA and AA-Briefcase come from the same evaluator and agree closely (r = 0.96), so the general-work part of the score leans on one method. The Finance and Legal indices both include Vals' Tax Agent test, a small overlap we accept. Muse Wonder Full has a Work Score but no public API, so it has no price or cost per task. Benchmarks are run by third parties on current model versions and can change when they update their tasks. A dash (–) means no score exists, never an estimated one. Prices come from Artificial Analysis and are standard API prices per 1M tokens.