Show the evidence. Explain the limits.
An AI score can tell you something useful. It cannot tell you everything.
Different tests measure different abilities. Some measure which answer people prefer. Others test maths, factual questions or software engineering. None of them tells you, on its own, which AI is best for every person.
In plain English
We read public test results and turn each into a simple position (best quarter, second, third or last) among the 77 models we list. We show where every number came from. When a test has no result, we say “Not enough data”. No AI picks the rankings, and nothing here is for sale.
A preference vote measures preference, not correctness. When people pick the answer they like better, they may favour one that is clearer, longer or more confident, including when it is wrong.
The data we use
Three public sources were approved. Licences, attribution, files and versions are on the data sources page.
| Source | Licence | Status here | Files used |
|---|---|---|---|
| Epoch AI Benchmark Hub | CC-BY-4.0 | Published | benchmark_metadata, frontiercode_external, frontiermath_tiers_1_3_v2, gpqa_diamond, model_metadata, otis_mock_aime_2024_2025, simpleqa_verified, swe_bench_verified |
| LMArena (Arena) leaderboard dataset | CC-BY-4.0 | Published | text, webdev |
| SWE-bench leaderboard (mini-SWE-agent runs) | CC-BY-NC-4.0 | Not published | leaderboards |
Where a source's overall licence is stated, that does not automatically settle the rights to every individual record. We say where that question is open.
What each test measures, and what it cannot tell you
Arena text: overall preference
- Version
- Arena 'text' leaderboard (Bradley-Terry rating, no style control)
- What it measures
- How often people chose this model's answer over another model's in blind, side-by-side chats on everyday prompts.
- What it cannot tell you
- Preference is not correctness. People can favour answers that look polished or confident, including answers that are wrong.
- Unit
- Arena score (people's votes)
- Models with a result
- 77 of 77 (rated)
Arena text: creative writing
- Version
- Arena 'text' leaderboard, category creative_writing
- What it measures
- People's blind votes on prompts classified as creative writing.
- What it cannot tell you
- It does not test whether facts are right, whether the tone suits your reader, or how well a model matches your own voice.
- Unit
- Arena score (people's votes)
- Models with a result
- 77 of 77 (rated)
Arena text: coding
- Version
- Arena 'text' leaderboard, category coding
- What it measures
- People's blind votes on prompts classified as coding questions.
- What it cannot tell you
- Voters judge answers by reading them. It does not prove the code runs, passes tests or is safe to ship.
- Unit
- Arena score (people's votes)
- Models with a result
- 77 of 77 (rated)
Arena text: maths
- Version
- Arena 'text' leaderboard, category math
- What it measures
- People's blind votes on prompts classified as maths.
- What it cannot tell you
- A vote for a maths answer is not a check that the answer is right.
- Unit
- Arena score (people's votes)
- Models with a result
- 77 of 77 (rated)
Arena text: following instructions
- Version
- Arena 'text' leaderboard, category instruction_following
- What it measures
- People's blind votes on prompts that ask for specific formats, lengths or constraints.
- What it cannot tell you
- It shows preference for how well instructions seemed to be followed, not an automatic check of each instruction.
- Unit
- Arena score (people's votes)
- Models with a result
- 77 of 77 (rated)
Arena text: business and management
- Version
- Arena 'text' leaderboard, category industry_business_and_management_and_financial_operations
- What it measures
- People's blind votes on prompts classified as business, management and financial operations.
- What it cannot tell you
- It is a broad topic label, not a test of emails, proposals or reports in your industry.
- Unit
- Arena score (people's votes)
- Models with a result
- 77 of 77 (rated)
Arena WebDev: building web apps
- Version
- Arena 'webdev' leaderboard (Code Arena), category overall
- What it measures
- People's blind votes on web pages and apps that each model built from the same request.
- What it cannot tell you
- It covers building small web projects. It says little about large existing codebases or other kinds of software.
- Unit
- Arena score (people's votes)
- Models with a result
- 61 of 77 (rated)
FrontierMath Tiers 1-3 (v2)
- Version
- FrontierMath Tiers 1-3 v2 (private set), benchmark released 2026-06-12
- What it measures
- The share of difficult, original maths problems with exact answers that the model solved when Epoch AI ran it.
- What it cannot tell you
- It measures hard problem solving. It does not show how well a model explains maths to a learner or checks your homework.
- Unit
- share of problems solved (measured accuracy)
- Models with a result
- 43 of 77 (rated)
OTIS Mock AIME 2024-2025
- Version
- OTIS Mock AIME 2024-2025, benchmark released 2024-12-19
- What it measures
- The share of AIME-style competition maths problems (from OTIS mock exams, 2024-2025) the model solved when Epoch AI ran it.
- What it cannot tell you
- Many leading models now score close to the maximum, so this test separates them poorly.
- Unit
- share of problems solved (measured accuracy)
- Models with a result
- 48 of 77 (rated)
GPQA Diamond
- Version
- GPQA Diamond, benchmark released 2023-11-20
- What it measures
- The share of graduate-level multiple-choice science questions answered correctly.
- What it cannot tell you
- Multiple-choice science exams do not show how a model copes with open-ended questions. Leading scores are close to the ceiling.
- Unit
- share of questions correct (measured accuracy)
- Models with a result
- 46 of 77 (rated)
SimpleQA Verified
- Version
- SimpleQA Verified, benchmark released 2025-09-09
- What it measures
- The share of short factual questions the model answered correctly in Epoch's runs.
- What it cannot tell you
- It tests recall of narrow facts under the test's own setup. It does not show how a model behaves with other tools, such as web search, or with questions that have no clear answer.
- Unit
- share of questions correct (measured accuracy)
- Models with a result
- 43 of 77 (rated)
SWE-bench Verified (Epoch runs)
- Version
- SWE-bench Verified, benchmark released 2024-08-13, as run by Epoch AI
- What it measures
- The share of real software bug reports the model fixed so that the project's own tests passed.
- What it cannot tell you
- It uses a fixed set of open-source Python projects. Results depend on the tooling around the model.
- Unit
- share of issues resolved (measured accuracy)
- Models with a result
- 11 of 77 (rated)
FrontierCode (reported by Cognition)
- Version
- FrontierCode, benchmark released 2026-06-08, aggregation Mean@5
- What it measures
- How well a model, used through a particular coding tool, completes difficult coding tasks, averaged over five attempts.
- What it cannot tell you
- Each row uses a particular coding tool (harness). Different tools make results not directly comparable between models, so we show these scores but never rate them.
- Unit
- mean score over 5 attempts (measured accuracy)
- Models with a result
- 34 of 77 (not rated)
How models get included
Which models appear is decided by a published rule, not by hand. A model is listed if it is in the top 60 on Arena's text vote (one entry per model: where Arena lists several settings of the same model, such as High and Max, we keep the best-ranked one) or it has results in at least three of the six Epoch benchmarks we use. The list is 77 models from 16 makers, and it is not weighted by maker. The rule is in tools/suggest_models.py and docs/MODEL_MAPPING.md.
An Epoch result is attached to a model only when its name matches exactly (after lower-casing, treating “.” and “-” alike) and Epoch's own family name for it agrees. Where it does not, that model keeps its Arena results and shows “Not enough data” for the Epoch tests. The rule has a limit: it favours models that Arena's voters and Epoch's runs already cover, so very new or niche models may be missing.
How we match models across sources
Each source names models differently. We keep a hand-edited list (config/models.json) with the exact identifier for each source. We never use fuzzy matching, and we never merge two versions or settings because their names look alike. A model whose identifiers have not been checked is held back and reported rather than published. 1 candidate is held back at the moment.
Sources also test models at different settings (for example “High” or “max” reasoning effort). We show the setting used for every result and do not adjust for it.
How ratings work
Ratings are relative. They compare only the 77 models on this site. Each of the four ratings (writing, coding, maths, knowledge) uses one named test. Models are sorted by their score, and the order is split into quarters.
| Band | Rule | Meaning |
|---|---|---|
| Top tier | position below 0.25 | In the best quarter of the models listed here. |
| Strong | position below 0.50 | In the second quarter. |
| Middle | position below 0.75 | In the third quarter. |
| Behind the leaders | position below 1.00 | In the last quarter of the models listed here. This is a relative position, not a verdict that the model is poor. |
A test is rated only when at least 8 listed models have a result. We do not combine unrelated tests into a single “intelligence score”. The four ratings are:
- Writing: Arena text: creative writing. How often people preferred its answers to creative writing prompts.
- Coding: Arena text: coding. How often people preferred its answers to coding prompts.
- Maths: FrontierMath Tiers 1-3 (v2). Share of hard maths problems it solved.
- Knowledge: SimpleQA Verified. Share of short factual questions it answered correctly.
Missing data
If a model has no result for a test, we show Not enough data. Missing is not zero, and a model is never marked down because it has fewer published results. Models without the main test for a task are listed as “not ranked” with the reason.
How recommendations work
Recommendations come from fixed rules in code and configuration. No AI model produces them, and the same data always gives the same result.
- Only models with a result for the task's main test, and at least min_tests results across the task's tests, are ranked. Everyone else is listed as 'not ranked' with the reason, never as worse.
- Eligible models are sorted by the main test's score, best first.
- Best-supported choice: the first eligible model whose supporting tests (where they have a result and a rating) do not include 'Behind the leaders'.
- Good-enough alternatives: the next eligible models, in score order, rated Top tier or Strong on the main test, with no supporting test rated 'Behind the leaders'. At most max_alternatives are shown.
- Best open-weights option: the highest-scoring eligible model that is not already shown, has an open licence (anything other than 'Proprietary' in Arena's data, or an open-weights label from Epoch AI), is rated Top tier or Strong on the main test and has no weak supporting test. It is a statement about the licence. It says nothing about price or free access.
- A model is flagged as a close call when its confidence interval (or +/-1.96 standard errors) overlaps with the best-supported choice's.
- Cheapest verified option: shown only when verified consumer-subscription prices exist (config/pricing.json) for at least one shown model. API prices are never used.
- No language model is used to produce rankings. The same inputs always give the same output.
Three labels mean three different things: Best-supported choice, Good-enough alternative and Cheapest verified option. The last appears only when verified consumer-subscription prices exist. We can compare performance, but we don't yet have enough verified pricing data to tell you which option is cheapest. We never use API prices as subscription prices.
When the data was updated
- Last successful update
- 29 Oct 2026, 14:36 UTC
- Data last changed
- 29 Oct 2026, 14:36 UTC
- Newest result observed
- 28 Oct 2026
- Oldest result observed
- 29 Jan 2026
- This page was generated
- 29 Oct 2026, 21:41 UTC
The update time is when the data was last checked against the sources and passed validation. The generation time is only when this page was built. They are different things. See what changed.
Exclusions and unresolved issues
- SWE-bench leaderboard (mini-SWE-agent): not published. The hosting repository's licence is NonCommercial and we cannot establish that this site's use is non-commercial in every future form. In addition, the mini-SWE-agent entries span agent versions 0.0.0 to 2.0.0 with different step and cost budgets, so they are not one uniform test. Affected records are excluded from every published number. They are kept in data/processed/swebench_audit.json for review.
- Rows republished from Artificial Analysis are excluded from Epoch data by checking source fields, not display names. None were found in the files used.
- FrontierCode is shown but never rated: its scores use different coding tools.
- Prices: there is no approved source for consumer subscription prices, so every price reads “Price not independently verified”.
The full list is on the data sources page.