Model classes S/A/B/C/D

🧠 Reasoning

Prices of these models in the catalog: class S and above · class A and above · class B and above

One source so far - score only

DeepSeek V4.1 Flash 1017 89.1, DeepSeek V4 Flash 0731 76.3, Kimi K2.7 Code 70.4, Qwen3.8 27B 63.6, DeepSeek V4 Flash 49.5, MiniMax-M2.7 47.6, MiniMax-M2.5 30.6

How we count

Every test result is put on a 0-100 scale against the strongest recent models. A source's score is the mean of its tests; the direction's score is the mean of its sources, so one source with many tests does not outweigh another. The strongest published variant of a model is used (e.g. the highest effort). Bands are fixed for half a year (edition 2026H2); the numbers are recomputed when a source publishes new results.

Class bands

• S from 95
• A from 80
• B from 62
• C from 42
• D from 0

Sources

• LMArena leaderboard dataset (CC BY 4.0) - blind human preference: math, expert prompts
• Epoch AI Benchmarking Hub (own runs) (CC BY 4.0) - GPQA Diamond, OTIS Mock AIME, FrontierMath

Next directions: code, agents, vision - one at a time, when each has two independent sources.

The same data: /v1/classes (API) and the MCP tool models_by_class.

A class says how a model scores on public tests, not whether it fits your job: try it on your own tasks.