FeatureBench: leaderboard
Metric: Resolved rate (%): share of tasks whose patched repository passes every fail-to-pass and pass-to-pass test, on the 200-task FeatureBench Full set (feature-level development tasks traced from unit tests in 24 Python repositories: 166 extending an existing codebase and 34 implementing the feature from scratch, each with a callable interface to implement, fail-to-pass and pass-to-pass tests, internet access without browser tools and cheating checks); each model run in the named agent harness (OpenHands up to 500 steps); higher is better. Source: arxiv.org. Saturation forecast: Around 2030. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.5 | 10.5 | #79 |
| 2 | DeepSeek V3.2 | 5.5 | #198 |
| 3 | Gemini 3 Pro (Preview) (Low) | 4.5 | #64 (Gemini 3 Pro (Preview)) |
| 4 | Qwen 3 Coder 480B A35B Instruct | 3.5 | #302 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=featurebench · How It Works · Data refreshed daily, snapshot 2026-10-11.