AISI Cyber TLO 10M: leaderboard
AISI cyber range: "The Last Ones", a 32-step corporate network attack simulation requiring reconnaissance, lateral movement, credential theft, reverse engineering, and data exfiltration. Avg steps completed at 10M token budget.
Metric: Avg Steps (/32). Source: arxiv.org. Saturation forecast: Around 2029. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 10 | #31 |
| 2 | Claude Opus 4.6 | 9.8 | #63 |
| 3 | Claude Mythos Preview | 9.5 | #4 |
| 4 | GPT-5.4 | 8.8 | #79 |
| 5 | GPT-5.1 Codex | 8 | #159 |
| 6 | Claude Opus 4.5 | 7.6 | #83 |
| 7 | GPT-5.3 Codex | 7.2 | #69 |
| 8 | Claude Sonnet 4.5 | 6.1 | #142 |
| 9 | Claude 3.7 Sonnet | 5.8 | #249 |
| 10 | GPT-4o | 1.7 | #343 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=aisi-cyber-tlo-10m · How It Works · Data refreshed daily, snapshot 2026-10-20.