METR, a non-profit research group that measures what advanced artificial intelligence systems can do on their own, said on Friday, 20 November 2026, that the most capable public AI models can now complete software and research tasks that would take a skilled human around 40 hours, with a 50% success rate. In its quarterly capability update, METR estimated that this "time horizon" is doubling roughly every 3.6 months, and concluded that "current public evidence is insufficient to rule out drastic AI R&D acceleration in the next couple of years."
What is METR's 'time horizon' measure?
The time horizon is the length of task, measured by how long it takes a skilled human, that an AI agent can complete successfully half the time. A model with a 40-hour horizon can, on average, finish about half of tasks that would take a human expert a full working week. MIT Technology Review described METR's chart of this measure as the most misunderstood graph in AI when reporting the update.
What did the new update find?
METR's main figures, for models publicly available up to Opus 5.5 and GPT-6 Astra, were:
- GPT-6 Astra: a 50% time horizon of 40 hours, with a very wide 95% confidence range of 13 to 2,645 hours.
- Opus 5.5: 39 hours, with a range of 12 to 1,253 hours.
- Doubling time: 3.61 months (about 110 days) across its combined task sets, close to the 105 days of its previous trend.
Under alternative estimates of how long humans take, the two leading figures fall to roughly 29 to 32 hours. METR said a simple extrapolation of the trend would reach its longest 200-hour measurement threshold in July 2027, though it treats such projections with caution.
The update also introduced new task sets, including 36 tasks in 12 families, four of which involve judgment and are hard to check automatically, such as writing new evaluation tasks for METR itself and scoring philosophical arguments.
Where are AI models still falling short?
On judgment-heavy work, the results were weaker. AI agents built evaluation tasks that METR rated "borderline-useful," but none were judged especially informative. On scoring philosophical arguments, many models beat experts who spent five minutes per item, but all public models stayed below experts who spent 15 minutes.
METR also found that AI judges were too generous: Opus 5.5 scored tasks written by GPT-6 Astra about 0.2 points (on a scale of 1) higher than a human scorer did. "By their nature, hard-to-verify or judgment-laden tasks are much harder to study, and especially to quantify," METR technical staff member Matt Putz told MIT Technology Review.
On specialised challenges, the strongest models placed in the 96th to 98th percentile of human participants in a set of mathematical optimisation contests. In a 48-hour attempt to speed up the training of a small language model, the best agent achieved a 15.7% speed-up, against a top human record of 41% reached after about 60 hours with AI help.
How reliable are these numbers?
METR was open about the uncertainty. About 91% of the human completion times are researchers' estimates rather than measured baselines, and the group called some of them "quite speculative." The confidence ranges for the top models span from about half a day to several months of human work. The human testers may also be less skilled than researchers at leading AI labs, and METR said it is unsure what capabilities would actually be needed for AI to drastically speed up AI research.
Why it matters
The question of whether AI systems can soon automate parts of AI research is central to debates on safety, regulation and the vast investments being made in the sector, which have come under pressure this week, including in the sell-off in Asian technology stocks on 20 November. METR's update suggests capabilities on well-defined tasks are still growing fast, while judgment remains harder both for the models and for those trying to measure them.




