Vendor-reported scores
AI Coding Leaderboard
Current models ranked on the coding benchmarks labs publish today. Each score comes from the provider's own release post or model card, and different versions of a benchmark are never mixed.
Agentic coding in a real terminal: building, debugging and running software end to end. The headline coding test in current frontier release posts.
| # | Model | Terminal-Bench 4.0 |
|---|---|---|
| 🥇 | Claude Sonnet 5.5 | 70.6%max effort |
| 🥈 | Claude Mythos 5.2Preview | 70.5%max effort |
| 🥉 | Claude Fable 5.2 | 67.6%max effort |
| #4 | Claude Opus 5.5 | 66.4%xhigh effort |
| #5 | Kimi K3.1 Pro | 65.7%max effort |
| #6 | Gemini 4 Argon | 57.4%highest thinking |
| #7 | Muse WonderPreview | 52.0%max effort |
| #8 | GPT-6.1 Terra | 50.3%max effort |
| #9 | ERNIE 5.5 Flash PreviewPreview | 46.5% |
| #10 | Hy4 | 44.4% |
| #11 | DeepSeek V4.1 Flash | 42.3%1017 update |
| #12 | Mistral Large 4 | 41.0% |
| #13 | Muse Ember | 40.9%max effort |
| #14 | Claude Haiku 5.5 | 39.2%max effort |
| #15 | MiniMax M3.1 Flash | 38.1% |
| #16 | Grok 4.7 | 37.6%xhigh effort |
| #17 | Step 5 PreviewPreview | 33.3%high reasoning |
| #18 | SWE-2 | 27.3% |
| #19 | SWE-2 Flash | 20.0% |
| #20 | DeepSeek V4 | 12.4%max effort |
No published Terminal-Bench 4.0 score
Scores are vendor-reported, so test setups differ between labs; text under a score is the setting the vendor used. Tilde (~) marks approximate figures. Models whose providers have not published any of these benchmarks are not ranked. Data last updated 2026-11-20. View full model specs →