Every Model Family · Tracked
One page per family · every Claude Opus variant, every Qwen 3 size, every Llama 4 release. Aggregate stats, price range, and the full lineage of each lab's work.
OpenAI
· 12 familiesAlibaba Qwen
· 10 familiesDeepSeek
· 4 familiesGoogle DeepMind
· 4 familiesstepfun
· 1 familyxAI
· 3 familiesminimax
· 2 familiesz-ai
· 4 familiesMicrosoft
· 2 familiesAnthropic
· 3 familiesMistral AI
· 4 familiesMeta
· 6 familiesCohere
· 1 familyNVIDIA
· 2 familiessao10k
· 2 familiesRed Hat AI
· 1 familyFrequently asked
How the hierarchy works
What is a model family?
A model family is a group of related AI model variants released by the same lab · for example Claude Opus ships as Opus 3, 3.5, 4, 4.6 and Qwen 3 ships as dozens of parameter sizes. BenchGecko tracks 65 families with 317 total variants.
Which family has the most variants?
Qwen 3 has 36 tracked variants, released by Alibaba.
Which family has the top-scoring variant?
The highest-scoring variant is GPT-5 Chat at 81.9 · part of the GPT-5 family by OpenAI.
Why not dedupe families into one canonical model?
Because per-variant pricing, benchmarks, and context windows differ materially. Deduping would destroy data quality. Instead we keep every variant in /models and expose the family as a second navigation layer · see ADR-0005.
See also
Keep exploring the BenchGecko graph