# MaxxBench (mutated DeepSWE tasks) (MaxxBench)

> unLAP's MaxxBench takes 89 DeepSWE software engineering tasks and changes the instructions or starting code while keeping the core problem, to test whether coding agents adapt to new requirements or fall back on a memorized solution.

Canonical page: https://benchlm.ai/benchmarks/maxxbench

- Category: [Coding](/coding)
- Last updated: November 7, 2026

## About MaxxBench

- Year: 2026
- Tasks: 89 mutated DeepSWE feature-implementation tasks, two attempts per model
- Format: Pass rate over 178 agent attempts, mini-swe-agent harness
- Difficulty: Long-horizon software engineering under changed requirements
- Paper: [Why MaxxBench (unLAP research post)](https://maxxbench.unlap.ai/research/why-maxxbench)

Each MaxxBench task starts from a DeepSWE feature-implementation task of medium length and changes what is asked (a renamed output, an extra constraint, a different interface) or the code the agent starts from, while the engineering problem stays the same. Reference patches grow by a median of 17 lines (3.5%). Every model runs in the same mini-swe-agent harness at the effort the source lists (max, xhigh or high), twice per task, and a run passes only if all required feature and regression tests pass in a separate grading container. The score is the share of the 178 attempts that pass. The leaderboard also shows an estimated benchmaxxing gap: the model's published DeepSWE score minus its MaxxBench score. Tasks, reference solutions, tests and per-run trajectories are published.

MaxxBench is currently displayed on BenchLM for reference, but it is excluded from the weighted scoring formula.

## Leaderboard (14 models)

| Rank | Model | Creator | Score |
|------|-------|---------|-------|
| 1 | [Claude Opus 5.5](/models/claude-opus-5-5) | Anthropic | 76.4% |
| 2 | [GPT-6 Astra](/models/gpt-6-astra) | OpenAI | 71.9% |
| 3 | [GPT-5.6 Sol](/models/gpt-5-6-sol) | OpenAI | 69.1% |
| 4 | [Claude Sonnet 5.5](/models/claude-sonnet-5-5) | Anthropic | 66.9% |
| 5 | [Claude Fable 5](/models/claude-fable) | Anthropic | 63.5% |
| 6 | [Claude Opus 5](/models/claude-opus-5) | Anthropic | 63.5% |
| 7 | [GLM-5.3](/models/glm-5-3) | Z.AI | 56.7% |
| 8 | [DeepSeek V4.1 Flash 1017](/models/deepseek-v4-1-flash) | DeepSeek | 55.6% |
| 9 | [GLM-5.3-Flash](/models/glm-5-3-flash) | Z.AI | 55.1% |
| 10 | [Gemini 3.8 Flash](/models/gemini-3-8-flash) | Google | 53.4% |
| 11 | [GPT-6 Luna](/models/gpt-6-luna) | OpenAI | 51.7% |
| 12 | [GPT-5.6 Luna](/models/gpt-5-6-luna) | OpenAI | 46.1% |
| 13 | [Muse Spark 1.2](/models/muse-spark-1-2) | Meta | 38.8% |
| 14 | [DeepSeek V4 Flash 0731](/models/deepseek-v4-flash-0731) | DeepSeek | 28.1% |

## FAQ

### What does MaxxBench measure?

Whether a coding agent can still solve a familiar DeepSWE task once its requirements or starting code change. The core engineering problem stays the same, so a model that really understands it should adapt; a model tuned on the original task may ship a solution that no longer fits.

### Which model leads MaxxBench?

Claude Opus 5.5 at max effort leads with 76.4% (136 of 178 attempts), ahead of GPT-6 Astra at xhigh (71.9%) and GPT-5.6 Sol at max (69.1%). Opus 5.5 is also the only model whose MaxxBench score is higher than its DeepSWE score.

### What is the benchmaxxing estimate?

The MaxxBench leaderboard subtracts each model's MaxxBench score from its published DeepSWE score. A large gap, such as 25.2 points for DeepSeek V4 Flash or 21.1 for GPT-5.6 Luna, suggests the model does much better on the original tasks than on close variants. It is an estimate, because the baselines come from different sources and settings.

## Compare Top Models on MaxxBench

- [Claude Opus 5.5 vs GPT-6 Astra](/compare/claude-opus-5-5-vs-gpt-6-astra)
- [GPT-6 Astra vs GPT-5.6 Sol](/compare/gpt-5-6-sol-vs-gpt-6-astra)
- [GPT-5.6 Sol vs Claude Sonnet 5.5](/compare/claude-sonnet-5-5-vs-gpt-5-6-sol)
- [Claude Sonnet 5.5 vs Claude Fable 5](/compare/claude-fable-vs-claude-sonnet-5-5)