代码与终端 agent 榜
Terminal-Bench 4.0：在真实容器里装环境、跑测试、读报错、改代码，所有模型跑同一脚手架。
| # | 模型 | 数值 | 相对第一 | 综合 v5 | 输出 $ | 类型 |
|---|---|---|---|---|---|---|
| 1 | Kimi K3.1 Pro 开源Moonshot AI | 67.2 | 第 1 | 46.0 | $10.00 | 开源 |
| 2 | Gemini 4 CarbonGoogle | 65.2 | 97% | 60.0 | $20.00 | |
| 3 | Claude Sonnet 5.5Anthropic | 63.6 | 95% | 54.5 | $10.00 | |
| 4 | Claude Opus 5.5Anthropic | 59.6 | 89% | 59.1 | $20.00 | |
| 5 | GPT-6 AstraOpenAI | 59.1 | 88% | 56.9 | $50.00 | |
| 6 | Gemini 4 ArgonGoogle | 57.1 | 85% | 50.4 | $10.00 | |
| 7 | GPT-6.1 SolOpenAI | 56.1 | 83% | 54.6 | $10.00 | |
| 8 | GPT-6.1 TerraOpenAI | 55.6 | 83% | 51.3 | $5.00 | |
| 9 | Claude Fable 5.1Anthropic | 52.0 | 77% | 49.0 | $50.00 | |
| 10 | GLM-5.3 开源Z.ai | 41.9 | 62% | 32.0 | $4.40 | 开源 |
| 11 | Qwen3.8 MaxAlibaba Qwen | 38.9 | 58% | 34.0 | $6.00 | |
| 12 | MiMo-V2.6 Pro 开源Xiaomi | 31.3 | 47% | 30.0 | $1.40 | 开源 |
| 13 | Grok 4.7xAI | 25.8 | 38% | 34.3 | $6.00 | |
| 14 | DeepSeek V4.1 Flash 开源DeepSeek | 19.7 | 29% | 33.0 | $1.20 | 开源 |
| 15 | Kimi K3 开源Moonshot AI | 17.2 | 26% | — | $15.00 | 开源 |
| 16 | DeepSeek V4 Pro 开源DeepSeek | 14.1 | 21% | 23.0 | $3.96 | 开源 |
| 17 | MiniMax M3 开源MiniMax | 1.0 | 1% | — | $1.20 | 开源 |