CFE-Bench: leaderboard
Metric: Question accuracy (%): share of the 449 CFE-Bench problems (305 text-only and 144 multimodal university STEM homework and exam problems with instructor-verified solutions) for which every annotated answer variable is correct; answers extracted and verified per annotated target variable by a GPT-5-mini judge (gpt-5-mini-2025-08-07) against expert-annotated values; chain-of-thought prompting, thinking capped at 16,000 tokens; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 59.69 | #54 |
| 2 | Gemini 3 Flash (Preview) | 55.46 | #78 |
| 3 | Gemini 3 Pro (Preview) | 55.01 | #64 |
| 4 | GPT-5.2 | 48.78 | #105 |
| 5 | Claude Opus 4.6 | 47.66 | #60 |
| 6 | Qwen 3.5 397B A17B | 47.44 | #141 |
| 7 | Qwen 3.5 Plus | 47 | #123 |
| 8 | Grok 4 | 41.65 | #169 |
| 9 | Grok 4.1 Fast (Reasoning) | 38.08 | #208 (Grok 4.1 Fast) |
| 10 | Llama 4 Maverick | 16.48 | #451 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=cfe-bench · How It Works · Data refreshed daily, snapshot 2026-10-11.