Timeline

METR says AI time horizons double every 3.6 months and acceleration cannot be ruled out

On a combined suite the doubling time is 3.61 months (95% CI 2.08 to 4.94); METR says its older evals may overstate real-world ability because they resemble RL training environments.

  • Benchmarks & progress
  • Safety & alignment
  • Notable

METR, the AI-evaluation nonprofit, published a capability update covering public models “up to Opus 5.5 and GPT-6 Astra”. It does not include Claude Fable 5.2 or GPT-6.2. METR’s headline is about uncertainty: it said current public evidence is “insufficient to rule out drastic AI R&D acceleration in the next couple of years”, defined as capabilities improving so quickly that humans can no longer meaningfully assess and react to catastrophic risks.

On its coding and optimisation tasks, METR said models usually match or beat skilled humans; when it recruited its own baseliners it typically failed to find any who outperformed the frontier models. On tasks needing choices without clear empirical feedback, it found the agents showed worse judgment than human experts. On the combined suite of its older TH1.1 tasks and the new, harder TH2 set, METR estimated time horizons doubling every 3.61 months (95% confidence interval 2.08 to 4.94), similar to the previous TH1.1 trend of 105 days, and noted that this resembles the roughly 3.8 months implied by OpenAI’s recently published data on its internal research tasks. On the verifiable subset of TH2 the doubling time was 1.65 months (interval 0.79 to 3.99). METR said this could reflect an acceleration hidden in the combined trend, or other causes, such as less diverse task sets, and that it was unsure which.

A second index, the METR Capability Index, showed no obvious acceleration or deceleration across all public frontier models tested. Fitted separately by developer it showed some slope increases, which METR said were comparable to its 5 November assessment of Anthropic and to Anthropic’s system cards, but sensitive to methodological choices. It listed reasons the method could miss a real acceleration, including developers delaying public releases.

The post is open about weaknesses in its own measurement. Within about a year of the March 2025 time-horizon release, METR said, models had saturated the TH1.1 suite. It also said it is “increasingly concerned” that easy-to-run evaluations resemble the reinforcement-learning environments developers train on, and so could overestimate real-world capability. Human baselines directly supplied only 9% of its task-duration estimates, and baseliners were allowed to use an AI model.

The update sets a third-party doubling-time estimate beside the lab-specific figure METR gave two weeks earlier, which put Anthropic’s rate of progress at roughly double its earlier pace. METR’s reading is that public-frontier data alone shows no clear acceleration, that it cannot exclude one, and that capabilities are getting harder to measure.

In the commentary

What people were saying around this time — external links, from the record's commentary rail.