Terminal-Bench 4.0
Terminal-agent tasks · 10 models · source Stanford / Laude Institute · updated 2026-10
What this benchmark measures
Software-engineering tasks inside a real container: install the environment, run the tests, read the errors, fix the code. Every model runs the same mini-SWE-agent scaffold so scaffolding is not a variable.
Why it gets its own table
This is the best available predictor of whether you can hand over a repository and walk away. The spread is enormous: 65% at the top, 20-40% in the middle, 1-2% for light open models.
Full ranking
| # | Model | Score | Vs. best | Input $ | Output $ | Type |
|---|---|---|---|---|---|---|
| 1 | Claude Opus 5.5Anthropic | 65.2 | Best overall | $4.00 | $20.00 | |
| 2 | Gemini 4 CarbonGoogle | 65.2 | 100% | $2.00 | $10.00 | |
| 3 | Claude Sonnet 5.5Anthropic | 64.0 | 98% | $2.00 | $10.00 | |
| 4 | GPT-6.1 TerraOpenAI | 55.6 | 85% | $1.00 | $5.00 | |
| 5 | GLM-5.3 Open weightsZ.ai | 38.9 | 60% | $0.60 | $2.20 | Open weights |
| 6 | Qwen3.8 MaxAlibaba Qwen | 34.3 | 53% | $2.00 | $8.00 | |
| 7 | MiMo-V2.6 Pro Open weightsXiaomi | 31.3 | 48% | $0.35 | $1.40 | Open weights |
| 8 | DeepSeek V4.1 Flash Open weightsDeepSeek | 19.7 | 30% | $0.30 | $1.20 | Open weights |
| 9 | Kimi K3 Open weightsMoonshot AI | 17.2 | 26% | $2.00 | $15.00 | Open weights |
| 10 | MiniMax M3 Open weightsMiniMax | 1.0 | 2% | $0.25 | $1.00 | Open weights |