More compute, more capability: Why AI agent evaluations need to account for test-time compute

AISI’s Science of Evaluation team published results indicating that standard practice of capping the compute an AI agent can use during evaluation underestimates its true capability. The study involved running frontier models at large test‑time budgets across a range of agentic benchmarks, including cybersecurity, software engineering, mathematics, academic tasks, and healthcare. Findings show that higher token budgets unlock harder tasks, with up to 8% of tasks only solved once the budget reached 10M+ tokens and some requiring up to 50M tokens. Performance gains from increased compute are larger for newer models, widening the gap between generations. The research also demonstrates that the compute required to solve a task scales with the time a skilled human would take, meaning fixed budgets cut off the longest tasks first. Consequently, fixed‑budget scores can misrepresent model comparisons, underestimate agent capabilities, and obscure risk scales. AISI’s analysis suggests that measuring capability curves over test‑time compute, rather than single scores, is necessary to accurately track progress and inform decision‑makers in the AI industry.