AISI Cyber TLO 100M: 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 100M token budget. Some scores are approximate (read from charts).

Metric: Avg Steps (/32). Source: www.aisi.gov.uk. Saturation forecast: Around January 2027. 8 models tracked.

Top models

#ModelScoreOverall rank
1GPT-5.522#31
2Claude Mythos Preview22#4
3Claude Opus 4.615.6#63
4GPT-5.414#79
5Claude Opus 4.511#83
6GPT-5.3 Codex11#69
7GPT-5.1 Codex9.8#159
8Claude Sonnet 4.59.4#142

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-100m · How It Works · Data refreshed daily, snapshot 2026-10-20.