MaxxBench: leaderboard
MaxxBench tests whether a coding agent follows a changed requirement instead of reproducing a solution it may have seen. It takes 89 feature-implementation tasks from DeepSWE and alters their instructions or starting code, with reference solutions and tests revised to match. Each model configuration attempts every task twice with mini-swe-agent in a sandboxed container at a stated reasoning effort, and a separate container grades the patch: an attempt passes only if every required feature and regression test passes. Unlike DeepSWE, its parent, the original task's known answer no longer satisfies the tests, so a model that scores lower here than there is leaning on the familiar task. This board is the share of the 178 attempts (89 tasks, two each) that pass.
Metric: Attempts passed on 89 changed DeepSWE tasks, two attempts each (%). Source: maxxbench.unlap.ai. Saturation forecast: Around March 2027. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5.5 (Max) | 76.4 | #5 (Claude Opus 5.5) |
| 2 | GPT-6 Astra (xHigh) | 71.91 | #6 (GPT-6 Astra) |
| 3 | GPT-5.6 Sol (Max) | 69.1 | #29 (GPT-5.6 Sol) |
| 4 | Claude Sonnet 5.5 (Max) | 66.85 | #18 (Claude Sonnet 5.5) |
| 5 | Claude Opus 5 (Max) | 63.48 | #20 (Claude Opus 5) |
| 6 | Claude Fable 5 (xHigh) | 63.48 | #26 (Claude Fable 5) |
| 7 | GLM-5.3 (Max) | 56.74 | #62 (GLM-5.3) |
| 8 | DeepSeek V4.1 Flash (Max) | 55.62 | #58 (DeepSeek V4.1 Flash) |
| 9 | GLM-5.3 Flash (Max) | 55.06 | #119 (GLM-5.3 Flash) |
| 10 | Gemini 3.8 Flash (High) | 53.37 | #36 (Gemini 3.8 Flash) |
| 11 | Muse Spark 1.3 (xHigh) | 53.37 | #42 (Muse Spark 1.3) |
| 12 | GPT-6 Luna (Max) | 51.69 | #93 (GPT-6 Luna) |
| 13 | GPT-5.6 Luna (Max) | 46.07 | #117 (GPT-5.6 Luna) |
| 14 | Muse Spark 1.2 (xHigh) | 38.76 | #97 (Muse Spark 1.2) |
| 15 | DeepSeek V4 Flash (Max) | 28.09 | #254 (DeepSeek V4 Flash) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #10 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=maxxbench · How It Works · Data refreshed daily, snapshot 2026-11-12.