SynCred-Bench: leaderboard
SynCred-Bench (Tsinghua CoAI, 2026) tests whether multimodal models prompted as detectors recognise AI-generated misinformation images that imitate credible forms such as notices, interfaces and certificates in realistic circulation styles. The score is the true positive rate in percent on the generated images at an operating point whose false positive rate on real images is at most 5 percent; higher is better.
Metric: True positive rate (%) on the 600 AI-generated credible-form images at an operating point whose false positive rate on the 450 real FP450 images is at most 5% (threshold swept over the stated confidence when the default FPR exceeds 5%); MLLM-as-judge detection prompt, image only. Source: arxiv.org. Saturation forecast: Around February 2027. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.6 | 69.5 | #102 |
| 2 | Claude Sonnet 4.6 | 55.3 | #161 |
| 3 | Gemini 3.1 Pro (Preview) | 12.7 | #94 |
| 4 | GPT-5.4 | 11.5 | #123 |
| 5 | Qwen 3.6 Plus | 3.7 | #222 |
| 6 | Qwen 3 VL 32B Instruct | 2.2 | #585 |
| 7 | Gemini 3.1 Flash Lite | 1.5 | #364 |
| 8 | GLM-5V Turbo | 0.5 | #344 |
| 9 | Grok 4.3 | 0.3 | #333 |
| 10 | GPT-4o | 0 | #741 |
| 11 | Qwen 3.5 9B | 0 | #803 |
| 12 | Qwen 3 VL 8B Instruct | 0 | #981 |
| 13 | Qwen 3.5 35B A3B | 0 | #482 |
| 14 | Llama 3.2 11B Instruct | 0 | #4991 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=syncred-bench · How It Works · Data refreshed daily, snapshot 2026-11-13.