METR estimates AI task horizons doubled roughly every 3.6 months

METR reported on November 20 that the task length frontier AI models can handle at a 50% success rate has roughly doubled every 3.6 months on two combined benchmark suites. Its evaluation covers publicly available models up to Anthropic’s Claude Opus 5.5 and OpenAI’s GPT-6 Astra.
METR measures a task’s length by how long it takes a human expert, not by how long an AI system runs. The group found that the harder of the two test suites lowered estimated task lengths at a 50% success rate compared with its earlier suite alone. Its separate index of overall model ability showed no strong evidence that the rate of improvement had accelerated or slowed across the public models it tested.
Models often matched or beat skilled people on METR’s coding and optimization tasks, but placed near the top in some comparisons with larger groups of specialists, but behind the strongest human performers. They also showed weaker judgment where feedback was unclear. The results show improvement on defined tests, not that agents can reliably handle a workweek without supervision.
Sources and citations
Revision history
Published .