AI & ComputeLOT 261122 · 3 min read
METR: Frontier AI Models Now Handle Tasks That Take Humans a Work Week
METR says frontier models including GPT-6 Astra now finish week-long human tasks, with task length doubling every 3.6 months — though judgment tasks remain out of reach.
Key takeaways
- METR said on Nov. 20 that frontier models including GPT-6 Astra complete tasks taking humans about one work week
- Task length frontier models can finish has doubled every 3.6 months, faster than the earlier four-month pace
- No AI-designed evaluation in METR's new task was good enough for staff to use; models fell short of humans at scoring philosophical arguments
- Anthropic's Opus 5.5 reached 'borderline above the bar' on evaluation-design tasks, up from 'clearly below' a year ago
- Most human completion times in the new suite are estimates; OpenAI said its models underperform on real-world employee tasks

By Nathan Brooks
Frontier AI models can now complete tasks that would take a human expert roughly one full work week, METR, the AI evaluation nonprofit, said on Friday, Nov. 20, according to The Information. The qualifying models include OpenAI's GPT-6 Astra, and the length of tasks frontier models can finish has continued to double every 3.6 months, METR reported.
That doubling figure comes from an update to METR's time-horizon chart — a graphic MIT Technology Review called widely shared and widely misinterpreted. The new version adds tasks estimated to take humans far longer than those in the original suite, along with more open-ended ones.
What does the time horizon actually measure?
The mechanism is simple. METR gives AI agents a suite of tasks that humans have also completed, then scores each task by how long a human expert took to finish it. A model's 50% time horizon is the task length at which it is predicted to succeed half the time.
The metric measures task difficulty, not how long a model can run unsupervised. On tasks they do complete, agents are typically several times faster than humans — speed was never the question. The earlier suites were mostly self-contained software engineering, machine learning and cybersecurity tasks with automatic scoring, and METR itself said measurements above 16 hours were unreliable with that setup.
The original chart, for context, showed the task length doubling about every four months, MIT Technology Review reported. The pace has quickened modestly, from four months to 3.6.
Where the models still fall short
METR added tasks closer to actual AI research work, and the results are instructive. In one, models designed new AI evaluations that METR employees then scored. None was good enough for staff to consider useful. In another, models scored philosophical arguments and fell short of human performance — though METR staff disagree on how relevant that task is to automating AI research.
The trend line is nonetheless moving. Some of the newest models, such as Anthropic's Opus 5.5, reached what METR calls "borderline above the bar" on the evaluation-design tasks, up from "clearly below the bar" for every model a year ago. Verdict: measurable progress on tasks nobody has automated yet.
How reliable are the numbers?
The report carries limits that METR openly acknowledges. Most human completion times in the new suite are estimates rather than measured times, because paying people to work through tasks of hundreds of hours is too costly. For the philosophy task, a single person rated most of the arguments.
"By their nature, hard-to-verify or judgment-laden tasks are much harder to study, and especially to quantify," METR's Matt Putz told MIT Technology Review.
OpenAI, for its part, said in a blog post on Wednesday that its models perform less well on real-world tasks assigned by its own employees than on the easier, verifiable tasks in METR's original suite, the magazine reported. That gap — between benchmark suite and workplace reality — is the one that matters commercially.
METR also evaluates models under the embedded safety-assessment arrangement OpenAI announced on Nov. 16.
The implication for buyers of AI tooling is straightforward: headline time-horizon figures will keep doubling on schedule, while performance on judgment-heavy work — the kind enterprises actually pay for — advances from "clearly below" to merely "borderline."
mlq.ai (Source)
More from Nathan Brooks
Show full bio
Senior reporter covering people, launches and day-to-day practice in Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.
1,201 articles