FireBench - Overconfidence: leaderboard
Metric: Share of the 370 overconfidence samples (%) in which the model performs the expected action: declining exactly the hard questions (GPQA Diamond, Humanity Last Exam, SimpleQA) the model answers wrongly under a standard prompt, and refusing unanswerable document questions; FireBench enterprise and API instruction following; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V3.1 Terminus | 86 | #212 |
| 2 | GPT-5.1 (Medium) | 67.7 | #131 (GPT-5.1) |
| 3 | Kimi K2 Instruct (0905) | 65.6 | #247 |
| 4 | GPT-5.1 Instant | 62.4 | #277 |
| 5 | Claude Sonnet 4.5 | 53.4 | #138 |
| 6 | Qwen 3 235B A22B 2507 Instruct | 44.2 | #291 |
| 7 | GPT-OSS-120B | 42.9 | #330 |
| 8 | Qwen 3 235B A22B 2507 (Thinking) | 42 | #253 (Qwen 3 235B A22B 2507) |
| 9 | GPT-4.1 | 38.6 | #240 |
| 10 | Llama 4 Maverick Instruct | 35.6 | #439 |
| 11 | Kimi K2 (Thinking) | 32.8 | #236 (Kimi K2) |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=firebench-overconfidence · How It Works · Data refreshed daily, snapshot 2026-10-11.