AlmanBench - Compliance: leaderboard
AlmanBench’s second metric: a conservative linter flags Standard German surface forms that Alman eliminates, and compliance is the share of responses that survive it. An unrelated but well-formed answer still counts as compliant, so this is a floor on quality rather than a measure of it — read it beside the acceptance board.
Metric: Compliance Rate (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 32 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 (xHigh) | 100 | #31 (GPT-5.5) |
| 2 | Kimi K3 (Max) | 100 | #30 (Kimi K3) |
| 3 | GPT-5.6 Sol (Max) | 100 | #19 (GPT-5.6 Sol) |
| 4 | Claude Opus 5 (Max) | 100 | #12 (Claude Opus 5) |
| 5 | DeepSeek V4 Pro | 99.9 | #102 |
| 6 | Claude Fable 5 (High) | 99.9 | #16 (Claude Fable 5) |
| 7 | Claude Fable 5 (Max) | 99.9 | #16 (Claude Fable 5) |
| 8 | Inkling (Max) | 99.9 | #126 (Inkling) |
| 9 | GLM-5.2 | 99.81 | #73 |
| 10 | GPT-5.6 Sol (xHigh) | 99.81 | #19 (GPT-5.6 Sol) |
| 11 | GPT-5.6 Terra (xHigh) | 99.81 | #41 (GPT-5.6 Terra) |
| 12 | DeepSeek V4.1 Flash (Max) | 99.71 | #36 (DeepSeek V4.1 Flash) |
| 13 | DeepSeek V4 Flash (0731) (Max) | 99.51 | #78 (DeepSeek V4 Flash (0731)) |
| 14 | Qwen 3.6 27B | 99.42 | #170 |
| 15 | Kimi K2.7 Code | 99.42 | #89 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=almanbench-compliance · How It Works · Data refreshed daily, snapshot 2026-10-20.