FireBench - Negative Content Requirements: leaderboard
Metric: Share of the 200 negative content requirements samples (%) in which the model performs the expected action: completing the same Arena-Hard 2.0 prompts while avoiding every forbidden operation, format or content type, judged by GPT-4.1; FireBench enterprise and API instruction following; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4.1 | 94.5 | #240 |
| 2 | Kimi K2 Instruct (0905) | 94.5 | #247 |
| 3 | Qwen 3 235B A22B 2507 Instruct | 93.5 | #291 |
| 4 | Llama 4 Maverick Instruct | 91.5 | #439 |
| 5 | Kimi K2 (Thinking) | 86 | #236 (Kimi K2) |
| 6 | Qwen 3 235B A22B 2507 (Thinking) | 84.2 | #253 (Qwen 3 235B A22B 2507) |
| 7 | GPT-5.1 (Medium) | 84 | #131 (GPT-5.1) |
| 8 | GPT-5.1 Instant | 82.5 | #277 |
| 9 | DeepSeek V3.1 Terminus | 81.4 | #212 |
| 10 | GPT-OSS-120B | 78.5 | #330 |
| 11 | Claude Sonnet 4.5 | 74.5 | #138 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=firebench-negative-content-requirements · How It Works · Data refreshed daily, snapshot 2026-10-11.