Blind-Spots-Bench - Text-only: leaderboard
Text questions that people find nearly trivial but AI models often fail, such as manipulating strings or satisfying hard counting constraints; from Blind-Spots-Bench (EPFL, 2026). Accuracy in percent averaged over four samples and graded against reference solutions; higher is better.
Metric: Accuracy (%; mean@4 over four samples on the text-only questions; 235 human-authored questions with reference solutions graded correct or incorrect by gemini-3-flash with code execution; no tools; thinking enabled at medium effort where available; max 32,768 output tokens). Source: arxiv.org. Saturation forecast: Around February 2027. 31 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 84 | #49 |
| 2 | Gemini 3.1 Pro (Preview) (Medium) | 83.3 | #94 (Gemini 3.1 Pro (Preview)) |
| 3 | GPT-5.4 (Medium) | 78.9 | #123 (GPT-5.4) |
| 4 | Gemini 3 Flash (Preview) (Medium) | 76.2 | #144 (Gemini 3 Flash (Preview)) |
| 5 | GLM-5.2 | 73.8 | #138 |
| 6 | Qwen 3.5 397B A17B | 71.1 | #246 |
| 7 | GPT-5 | 70.6 | #167 |
| 8 | GPT-5.2 (Medium) | 70.6 | #176 (GPT-5.2) |
| 9 | DeepSeek V4 Pro (Reasoning) | 70.4 | #188 (DeepSeek V4 Pro) |
| 10 | Kimi K2.5 | 68.3 | #256 |
| 11 | GPT-5 Mini | 67.6 | #348 |
| 12 | DeepSeek V4 Flash (Reasoning) | 67.6 | #258 (DeepSeek V4 Flash) |
| 13 | GLM-5.1 | 67.4 | #179 |
| 14 | Kimi K2.6 | 65 | #178 |
| 15 | GLM-5 | 64.4 | #260 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=blind-spots-bench-text-only · How It Works · Data refreshed daily, snapshot 2026-11-13.