Model classes S, A, B, C, D: how AI Rate Card grades reasoning models

Updated Nov 7, 2026

Prices tell you what a model costs; they do not tell you how strong it is. So base models on AI Rate Card now carry a class: one letter per direction, S at the top and D at the bottom. The first direction is reasoning - maths, science and expert questions. Code, agents and vision follow one at a time.

Where you see it

On a model's page, next to its id: 🧠 A. Hover it (or tap it on a phone) for the score and each source's score.
On the Model classes page: every graded model, best first, with the class bands and the sources.
In the API: GET /v1/classes?direction=reasoning&min_class=A, and a classes field in GET /v1/models/{model_id}.
Over MCP: the tool models_by_class (direction, min_class) on https://airatecard.com/mcp.

How a class is counted

Two independent sources. LMArena (blind votes of people on maths and expert prompts) and Epoch AI's own benchmark runs (GPQA Diamond, OTIS Mock AIME, FrontierMath). Both publish their data under CC BY 4.0.
One scale. Each test result is put on 0-100 against the strongest recent models; the strongest published variant of a model is used (for example its highest effort setting).
One vote per source. A source's score is the mean of its tests, and the model's score is the mean of the sources - so a source with many tests does not outweigh the other.
Fixed bands. For reasoning in edition 2026H2: S from 95, A from 80, B from 62, C from 42, D below. The bands stay put for half a year, so a letter does not move just because the scale moved.

A model measured by only one source gets its score and no letter. When the two sources disagree by two classes or more, the model is marked under review: both scores are on the page, so you can see the split yourself.

What the first grading shows

Of 49 reasoning models: 5 in S, 15 in A, 12 in B, 6 in C, 4 in D, and 7 with a score but no letter yet. The S class today is Claude Fable 5.1, Claude Sonnet 5.5, Claude Opus 5, Claude Opus 5.5 and GPT-5.6 Sol. GPT-6 Astra lands in A at 93.7 - Epoch gives it 100, LMArena voters 87.5; that is a real split between the sources, not a data error.

How to use it

Shortlist, then test. Take the models of the class you need, then try two or three of them on your own prompts: a class says how a model scores on public tests, not whether it fits your job.
Look for a cheaper model of the same class. An A-class model can cost a fraction of another A-class one; compare them on /compare or price your monthly volume on /tools.
Check the split. A letter under review is a hint that people and benchmarks see the model differently.

The data behind every letter is on the Model classes page and in /v1/classes; scores are recomputed when a source publishes new results.