Benchmarks
351 sources · 1,261 models × 3,053 benchmarks · 160,189 scoresPerfArena (NYU Abu Dhabi and the Universidade do Algarve) has each model write one solution per problem and language for 99 LeetCode problems in seven languages, and times every accepted solution on one bare-metal host. The score is a Bradley–Terry win probability from latency duels on the cells both models solved, so 50 is the cohort average and values shift as models are added. It is shown as a diagnostic outside the capability fit.
No signal or difficulty estimate. Held out of the fit by design.
| # | Model | Setting | Score | Source |
|---|---|---|---|---|
| 1 | GPT-6 Astra | not reported | 62.0 | |
| 2 | Claude Opus 5.5 | not reported | 58.0 | |
| 3 | GPT-5.6 Sol | not reported | 56.0 | |
| 4 | Claude Sonnet 5.5 | not reported | 54.0 | |
| 5 | Claude Fable 5.1 | not reported | 53.0 | |
| 6 | GLM-5.3 Flash | not reported | 53.0 | |
| 7 | Gemini 3.8 Flash | not reported | 51.0 | |
| 8 | GPT-5.5 | not reported | 49.0 | |
| 9 | GPT-5.4 Mini | not reported | 47.0 | |
| 10 | GPT-5.6 Luna | not reported | 46.0 | |
| 11 | DeepSeek V4.1 Flash | not reported | 43.0 | |
| 12 | Opus 4.8 | not reported | 39.0 | |
| 13 | Grok 4.6 | not reported | 39.0 |
Benchmarks outside its family whose model rankings track it most reliably (Spearman ρ over at least 10 shared models).