Industry benchmarks are systematically underestimating the true potential of top-tier AI agents by artificially restricting their "thinking time." According to a report from the UK AI Safety Institute (AISI), the standard practice of fixing strict computational budgets during testing provides a distorted picture. When models are allowed to utilize more resources, success rates in complex tasks jump by 25%. For business leaders, this represents a critical diagnostic error: you aren't measuring the system's ceiling, but rather its performance under a forced resource deficit.
The Power Law of Computation
AISI researchers found that effectiveness in cybersecurity and software development correlates directly with the compute budget. In cyber-exploitation tests, about 8% of tasks were only solved after the limit exceeded 10 million tokens, with some requiring up to 50 million. An agent's performance is not a single point on a graph; it is a curve that climbs as test-time compute increases. If you cut the budget while this curve is still rising, the final score reflects your frugality rather than the model's limitations. On the TerminalBench 2.0 and SWE-Bench Pro benchmarks, success rates surged by a quarter immediately after expanding limits from one million to ten million tokens.
Dependency follows a power law tied to human labor hours. Data from METR and AISI show that the volume of tokens an agent needs scales proportionally to the time a human expert would require. A one-minute task costs thousands of tokens; an hour-long task costs millions; a week-long project can consume billions.
Current evaluation methods effectively cut off the most complex logical nodes. If your internal pilots are stalling on difficult tasks, the problem might be tightened belts rather than the inherent intelligence of Claude 3.5 or o1 models.
The Efficiency Trap and Real TCO
Additional computation isn't a panacea, but its effect is predictable. In medical benchmarks like HealthBench, models hit a plateau regardless of budget. AISI attributes this to the environment: "extra" tokens add value where an agent can verify its own work—such as running code or testing an exploit. Where feedback is absent or subjective, there is almost no progress. This creates a strategic gap: heavy-duty agents are ready for autonomous technical processes but remain ineffective in "fuzzy" qualitative environments.
It is time to recalculate the Total Cost of Ownership (TCO) of your AI pilots by shifting from "price per thousand tokens" to "Cost per Task Success."
If a tenfold increase in token limits yields a 25% gain in a critical development bottleneck, that expense will almost certainly result in net savings on expensive human oversight. As token costs continue to plummet, capabilities that seemed economically absurd yesterday are becoming the standard today, turning yesterday's benchmark results into scrap paper.