AI model coding benchmarks
Every coding benchmark score found in our tool pages and in the model comparison, grouped by test. Each score comes with its stated conditions and a link to the vendor's page.
DeepSWE v1.1
| Sources | ||||
|---|---|---|---|---|
| Gemini 4 Argon | 77.9% | announced by Google | source — Gemini 4 Argon | |
| DeepSeek V4.1-Flash-1017 | DeepSeek | 77% | announced by DeepSeek, Mini-SWE-V2 harness | source — DeepSeek V4.1-Flash-1017 |
| MiniMax M3.1 Flash | MiniMax | 73.7% | announced by MiniMax | source — MiniMax M3.1 Flash |
| Grok 4.7 | xAI | 71% | high effort, announced by xAI | source — Grok 4.7 |
| Kimi K3 | Moonshot AI | 67.3% | mini-SWE-agent harness (DeepSWE leaderboard), cited by Moonshot | source — Kimi K3 |
| GLM-5.3 | Z.ai | 66.9% | announced by Z.ai | source — GLM-5.3 |
| GLM-5.3-Flash | Z.ai | 63.4% | announced by Z.ai | source — GLM-5.3-Flash |
| DeepSeek V4-Pro-0813 | DeepSeek | 62.7% | DeepSeek Harness (minimal mode), max effort, announced by DeepSeek | source — DeepSeek V4-Pro-0813 |
| Muse Spark Open (816B-A42B) | Meta | 60% | announced by Meta, preferred harness, xhigh reasoning | source — Muse Spark Open (816B-A42B) |
| Qwen3.8-Max | Alibaba (Qwen) | 56.6% | best of Claude Code / mini-SWE-agent, announced by Qwen | source — Qwen3.8-Max |
| Qwen3.8-27B | Alibaba (Qwen) | 42.2% | Claude Code harness, announced by Qwen | source — Qwen3.8-27B |
Terminal-Bench 4.0
| Sources | ||||
|---|---|---|---|---|
| Kimi K3.1 Pro | Moonshot AI | 65.7% | Claude Code harness, reported by Moonshot | source — Kimi K3.1 Pro |
| Gemini 4 Carbon | 65.2% | Artificial Analysis harness, reported by Google | source — Gemini 4 Carbon | |
| GPT-6.1 Terra | OpenAI | 50.3% | max reasoning effort, reported by OpenAI | source — GPT-6.1 Terra |
| Nanbeige4.3-33B | Nanbeige (BOSS Zhipin) | 23.8% | mini-swe-agent v2.4.6 harness, reported by Nanbeige | source — Nanbeige4.3-33B |
Model comparison → Price / performance chart → Cost calculator →