ResearchNews
METR: AI agent time horizons double every 3.6 months
METR's quarterly update finds frontier-model time horizons doubling about every 3.6 months. On coding tasks, models now usually match or beat skilled humans, but their judgment still lags behind experts.
· 3 min read

What did METR measure?
METR, the independent evaluation nonprofit, published its Q4 2026 update on how much AI models could accelerate AI research on November 20 (METR). It covers public models up to Opus 5.5 and GPT-6 Astra. A time horizon is the length of task, measured by how long it takes a human expert, that a model completes with 50% reliability (METR: time horizons).
The headline numbers
- Across the older TH1.1 and the new, harder TH2 suites, time horizons double roughly every 3.61 months (110 days), with a 95% CI of 2.08 to 4.94 months. That is close to the previous 105-day trend.
- Naive extrapolation puts models at the suite's longest 200-hour threshold in July 2027.
- The doubling time is shorter on TH2's verifiable subset, which could hint at a recent acceleration, though METR notes other explanations.
- In Al Zimmermann's months-long programming contests, the strongest models (Fable 5.1, Opus 5.5, GPT-6 Astra) placed at a median of the 96th to 98th percentile of human participants.
Strengths and weaknesses
On METR's coding and optimization tasks, models usually match or outperform skilled humans. When METR recruited its own baseliners, it often could not find people who beat the frontier models. Where many domain specialists had attempted the same task, models placed at a very high percentile, but not at the very top.
Two weak spots stood out. Models do worse on tasks that would be costlier to turn into RL training environments, although they improve at a similar rate on those. METR estimates that doubling a task's "lifecycle cost" roughly halves an agent's time horizon. And on tasks without clear empirical feedback, agents show worse judgment than human experts.
Why it matters
METR's main conclusion is that current public evidence cannot rule out drastic AI R&D acceleration in the next couple of years, meaning a pace at which humans can no longer meaningfully assess the risks. Measuring that is also getting harder as models outgrow the task suites. It lines up with OpenAI's data from last week showing agents now handle half of 2 to 4 hour research tasks (OpenAI: agents now handle half of 2-4h research tasks).
The practical reading: in coding and optimization, agents can now take on serious multi-hour tasks, but where a call needs taste or judgment without feedback, human review is still the bottleneck.
Produced with the help of AI, reviewed by an editor. How we make our articles
Reviewed and edited by the Neural Daily desk. About


