InfoOps Bench: leaderboard
Each item takes a claim from tracked Russian, Chinese and Iranian information operations and asks for a shareable social media post about it, in four framings such as an influencer persona. A classifier model sorts each reply, and replies that preserve or amplify the claim count against it. The integrity score is the share of untrue high harm claims not preserved or amplified, higher is better; claims update weekly.
Metric: Integrity Score (% refused). Source: pattrn.ai. Saturation forecast: Around February 2027. 46 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.8 | 97 | #39 |
| 2 | Claude Opus 4.7 | 95.7 | #51 |
| 3 | Claude Sonnet 5 | 93.3 | #65 |
| 4 | Claude Opus 5 | 92.3 | #16 |
| 5 | Claude Haiku 4.5 | 91.6 | #287 |
| 6 | GPT-6 Sol | 88.5 | #19 |
| 7 | Claude Haiku 5.5 | 87.6 | #43 |
| 8 | Claude Sonnet 5.5 | 87.5 | #14 |
| 9 | GPT-6 Luna | 86.2 | #61 |
| 10 | Claude Opus 5.5 | 85.7 | #3 |
| 11 | GPT-6.1 Sol | 85.1 | #7 |
| 12 | Claude Sonnet 4.6 | 84.4 | #94 |
| 13 | Qwen 3.5 Flash | 83.9 | #237 |
| 14 | Kimi K3 | 83.6 | #35 |
| 15 | Claude Opus 4.6 | 80 | #67 |
No result here: #6 GPT-6 Astra, #10 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=infoops-bench · How It Works · Data refreshed daily, snapshot 2026-11-05.