UK AISI measurements on its task suite; the size of the effect varies by model and task. Anthropic's September 2026 Opus 5.5 card reports ExploitGym exploits within two-hour and six-hour wall-clock budgets (Opus 5.5: 289 and 300; Mythos 5.1: 203 and 264): a longer budget added little for one model and more for the other, in wall-clock time rather than the token caps above.
Reported: Stated by one source and not yet corroborated or challenged.
Evidence
Raising the budget from 2.5M to 50M tokens moved one model's cyber time horizon from about 40 minutes to about 4 hours.
UK AISI's May post already said the 2.5M-token cap understates frontier capability.
On ExploitGym, Anthropic reports Claude Opus 5.5 at 289 exploits within two hours and 300 within six (Mythos 5.1: 203 and 264), so extra wall-clock time added little for Opus 5.5 and more for Mythos 5.1.
Project Zero’s BigSleep rediscovery rate for historical V8 bugs is reported over 25 attempts per introducing commit, so the figure depends on the attempt budget.
How it relates to other findings
Select a finding to see how it relates to others. Arrows point from the newer finding to the one it supports, qualifies, contests, or supersedes.
- This finding qualifies Under a 2.5M-token cap, frontier cyber task time horizons doubled on the order of months between late 2024 and early 2026.The doubling estimate was measured at a fixed low budget.
Key questions that rely on this finding
- How far can measured AI cyber capability be trusted?As a lower or conditional bound. Scores move with budget, pipeline and contamination, and benchmark names or labels can hide different setups.
Status history
- 2026-05-29ReportedOpenAI's evaluation playbook warns that unreported budgets understate capability. · record
- 2026-07-02CorroboratedUK AISI measures the effect directly. · record
- 2026-09-25ReportedcorrectionThe OpenAI playbook cites UK AISI's own measurements, so all evidence comes from one evaluator. · record