HELM VHELM - Blink Knowledge - Forensic Detection: leaderboard
VHELM (Stanford CRFM HELM for vision-language models, run v2.1.2) BLINK Forensic Detection task: which image is a real photograph rather than a generated one, one of BLINK's perception tasks that humans solve at a glance, asked as multiple choice; 132 questions, prefix quasi-exact match. Percent correct, 0-100, higher is better.
Metric: Prefix Quasi-Exact Match (%). Source: crfm.stanford.edu. Saturation forecast: Estimated already saturated. 48 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4.5 | 97.73 | #278 |
| 2 | GPT-4.1 Mini | 96.97 | #363 |
| 3 | O1 (2024-12-17) | 96.97 | #157 |
| 4 | Gemini 2.5 Pro (Preview 03-25) | 93.18 | #126 |
| 5 | GPT-4.1 (2025-04-14) | 91.67 | #257 |
| 6 | GPT-4o (2024-05-13) | 91.67 | #337 |
| 7 | O4 Mini (2025-04-16) | 90.15 | #183 |
| 8 | O3 (2025-04-16) | 88.64 | #125 |
| 9 | Gemini 2.0 Flash (Preview) | 88.64 | #376 |
| 10 | GPT-4o (2024-11-20) | 87.88 | #392 |
| 11 | Gemini 2.0 Flash | 86.36 | #351 |
| 12 | GPT-4o (2024-08-06) | 86.36 | #343 |
| 13 | Gemini 2.0 Pro (Preview 02-05) | 86.36 | #318 |
| 14 | Gemini 1.5 Pro (002) | 81.82 | #368 |
| 15 | Llama 4 Maverick Instruct FP8 | 76.52 | #424 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=helm-vhelm-blink-knowledge-forensic-detection · How It Works · Data refreshed daily, snapshot 2026-11-05.