Does the test resemble the intended work?
Does the test really resemble the work the tool will have to do?
Understanding AI at work / Short page
Understand · Short page
AI benchmarks: how to read a model leaderboard
A leaderboard says which model succeeds at a given test, not which one suits your work. This page summarises the families of tests, what each leaves out and the questions to ask before drawing a conclusion.
The essentials
There is no universal leaderboard.
A score describes a test, not your work.
Direct answer
There is no universal ranking of the “best model”. A benchmark measures performance on a set of tasks, with a given prompt, metric and conditions. Change one of these and the ranking can change.
None of these tests says which model suits your use: that is checked on your own tasks.
01 / Families of tests
Five families of tests, five limits.
Knowing what a test measures also means knowing what it leaves out.
| Family | Principle and examples | Main limit |
|---|---|---|
| Knowledge | Multiple-choice questions, marked automatically. Examples: MMLU, GPQA. | Highly exposed to contamination: the questions end up in the training data. |
| Real tasks | The model must produce a verifiable result: fix a program, deliver a project. Examples: SWE-bench, Remote Labor Index. | Costly to build and to score; they cover only part of a job. |
| Human preference | Two anonymous answers, one person votes. Example: LMArena. | Measures appeal as much as accuracy. |
| Autonomous behaviour | The model acts over time, with no expected answer at each step. Examples: METR’s time horizon, Vending-Bench. | No single “right answer”; the result depends heavily on the test set-up. |
| Occupation | A test built by professionals in the field. Examples: MedHELM in medicine; BenchMedTrav in occupational health, designed by the author of this site. | Closest to the job, but real deployment still has to be assessed. |
02 / Dated snapshot
The leader changes with the test.
As of 13 November 2026, and soon out of date: follow the links for current results.
| Test | What it measures | Leader |
|---|---|---|
| Capabilities Index (ECI) ↗ · Epoch AI | An aggregate of several dozen tests. | Claude Fable 5.2 |
| GDPval-AA ↗ · Artificial Analysis | Professional tasks, compared in pairs. | Claude Opus 5.5 |
| ARC-AGI-3 ↗ · ARC Prize | Adaptation to novel environments, under the standard protocol. | Claude Fable 5.2, with 81.2% |
Two precautions
The gaps between the leaders are often smaller than the measurement uncertainty. And the score depends on the protocol: on ARC-AGI-3, with each provider’s own set-up, four models exceed 97%.
03 / Method
Four questions before drawing a conclusion.
To ask of any score, whatever the test.
Under what conditions?
Which prompt, which tools, how many attempts and which model version?
What does the score capture?
Does the score measure accuracy, preference, cost, speed or something else?
What does the average hide?
Could the data have been seen during training, and which errors are masked by the average?
04 / At work
From score to use: two different assessments.
A good technical score does not, by itself, predict good professional use.
Does the system answer correctly?
Accuracy on representative cases, frequent errors and serious errors, stability from one answer to the next, cost and delay, resistance to misleading inputs.
Is the work done better without creating new risks?
Time actually saved, review included; final quality and handling of exceptions; responsibilities and means of appeal; workload, autonomy, skills and cooperation; effects observed before and after.
Perceived gain, measured gain
In a randomised trial run by METR in early 2025 with 16 experienced developers, they believed they were 20% faster with AI; they were measured 19% slower. A new measurement in late 2025 suggests a speed-up instead, which METR considers weak evidence. Becker, Rush, Barnes et al., METR, 2025 ↗
Key answers
Three frequent questions.
Short answers, drawn from the guide.
Which is the best AI model?
There is no single answer: the leader changes with the test, and the gaps between the leaders are often smaller than the measurement uncertainty. The useful question is which model succeeds at your own tasks.
What is benchmark contamination?
It is when a test’s questions end up in the models’ training data. Multiple-choice knowledge tests are highly exposed to it: before drawing a conclusion, ask whether the data could have been seen during training.
Are there benchmarks for occupational health?
Tests built by professionals in the field are the closest to the job: MedHELM in medicine, or BenchMedTrav in occupational health, designed by the author of this site. Even closest to the job, real deployment still has to be assessed.
Where this page comes from
It summarises one chapter of the guide Understanding AI at work, which gives the source, date and limits of every figure. The figures dated November 2026 will go out of date: the guide is the version kept current.
Update log
- First version, drawn from the guide “Understanding AI at work” updated on 13 November 2026.