CFE-Bench (Text) - Variable Accuracy: leaderboard
Metric: Variable accuracy (%): mean over the 305 text-only CFE-Bench problems (university STEM homework and exam problems with instructor-verified solutions) of the share of annotated answer variables answered correctly; 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. 26 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 70.66 | #54 |
| 2 | Gemini 3 Flash (Preview) | 66.02 | #78 |
| 3 | Gemini 3 Pro (Preview) | 65.29 | #64 |
| 4 | Claude Opus 4.6 | 58.95 | #60 |
| 5 | GPT-5.2 | 57.99 | #105 |
| 6 | Qwen 3.5 Plus | 54.43 | #123 |
| 7 | Qwen 3.5 397B A17B | 54.12 | #141 |
| 8 | Grok 4 | 53.24 | #169 |
| 9 | Kimi K2.5 | 51.32 | #139 |
| 10 | DeepSeek V3.2 (Thinking) | 50.07 | #198 (DeepSeek V3.2) |
| 11 | Grok 4.1 Fast (Reasoning) | 49.58 | #208 (Grok 4.1 Fast) |
| 12 | Claude Opus 4.5 | 49.03 | #79 |
| 13 | DeepSeek V3.2 (Non-reasoning) | 48.08 | #198 (DeepSeek V3.2) |
| 14 | GLM-5 | 47.24 | #137 |
| 15 | Kimi K2 (Thinking) | 46.28 | #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=cfe-bench-text-variable-accuracy · How It Works · Data refreshed daily, snapshot 2026-10-11.