Featherbench: leaderboard
Headline pass rate across 28 native Featherbench tasks: 22 coding, data, everyday or tool tasks and six security tasks. The website uses all 28 trials, including blocked/refused outcomes; this differs from the evaluation script’s scored-only pass denominator. Some security tasks explicitly pass on refusal. The published model rows combine run dates and checker revisions; four non-security checks changed after August 14 and 24 task blobs are unchanged. Native percentages are retained without recomputing or pooling the separate Rubric Quality score.
Metric: Pass Rate (%). Source: reinvently.co.uk. Saturation forecast: Estimated already saturated. 34 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V4.1 Flash | 100 | #38 |
| 2 | Claude Opus 5.5 | 100 | #3 |
| 3 | GLM-5.3 | 100 | #45 |
| 4 | Gemini 4 Argon | 100 | #9 |
| 5 | Claude Sonnet 5.5 | 100 | #13 |
| 6 | Grok 4.7 | 100 | #26 |
| 7 | Claude Haiku 5.5 | 100 | #46 |
| 8 | Kimi K3.1 Pro | 100 | #18 |
| 9 | Beam 1.0 | 100 | #107 |
| 10 | Muse Spark Open | 100 | #41 |
| 11 | Muse Wonder | 100 | #16 |
| 12 | GLM-5.3 Flash | 99 | #79 |
| 13 | Claude Sonnet 4.6 | 96 | #97 |
| 14 | GPT-5.5 | 96 | #34 |
| 15 | Claude Haiku 4.5 | 96 | #291 |
Interactive version: theaggregate.ai/benchmark?slug=featherbench · How It Works · Data refreshed daily, snapshot 2026-11-07.