RiskChainBench - End-to-End: leaderboard
RiskChainBench entry-gated end-to-end accuracy: the same 600 websites, but a failure to restore the obfuscated entry message or to route correctly also counts as wrong. Higher is better. Source: arxiv.org/abs/2609.16900.
Metric: Entry-gated Accuracy (%). Source: arxiv.org. Saturation forecast: Around February 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Sol | 60.8 | #22 |
| 2 | GPT-5.4 | 52.3 | #83 |
| 3 | Claude Opus 4.8 | 49.2 | #39 |
| 4 | GPT-5.2 | 49 | #115 |
| 5 | Kimi K3 | 44.5 | #35 |
| 6 | Kimi K2.5 | 25.7 | #152 |
| 7 | Qwen 3.6 Plus | 19.3 | #124 |
| 8 | Claude Sonnet 5 | 19.2 | #65 |
| 9 | Kimi K2.6 | 16.7 | #113 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=riskchainbench-end-to-end · How It Works · Data refreshed daily, snapshot 2026-11-05.