METR posted a new number on May 8, 2026.4 The nonprofit times how long an AI agent can work alone. An early version of Anthropic's Claude Mythos Preview finished half its hardest tasks. It worked alone for 16 straight hours first.
That score broke METR's own test. Only 5 of its 228 tasks run that long, so the result carries a wide range: 8.5 to 55 hours.4 Product leaders now face a real question. Does a half-day of unsupervised work mean an agent can finally run a real job alone?
The Clock Is Doubling Every Three Months
METR's own trend line explains why Mythos surprised everyone. The nonprofit calls its benchmark Time Horizon. It fits a doubling rate to every frontier model it tests.1 Models released since 2019 double their clean-task score every 196.5 days, about six and a half months.
Counting only models since 2023, the rate drops to 130.8 days. Since 2024, that pace nearly doubled again: the rate fell to 88.6 days, near three months.
AI Agents' Doubling Time Shrinks From 196 to 89 Days
Days for an AI agent's 50 percent success horizon to double, by models released since each start year, METR's Time Horizon 1.1
Each point is the doubling time METR calculated using only models released since that year; the line falls because recent models double their task window faster than older ones did.
Source: METR [1]. Chart by Kelly Dennis.
The numbers behind this chart
| Period | Value |
|---|---|
| Since 2019 | 196.5 days |
| Since 2023 | 130.8 days |
| Since 2024 | 88.6 days |
The climb looks steep on a chart.1 Claude 3.7 Sonnet needed one hour to clear half its tasks on METR's suite. GPT-5 reached 3.6 hours, or three hours and thirty-four minutes. Claude Opus 4.5 hit 5.3 hours, or five hours and twenty minutes, the best score before Mythos.
Mythos broke the scale on May 8, 2026.4 Its early version needed sixteen hours to clear the same bar, three times Opus 4.5's score.
Claude Mythos's Task Window Hits 16 Hours, Up From 1 in 2025
METR's 50 percent success horizon on its evaluation suite, by model, hours
Each column is one model's 50 percent success horizon in hours; the Mythos column is the tallest and the first to run past METR's reliable measuring range.
Mythos's score carries a wide confidence interval, and only 5 of METR's 228 tasks run 16 hours or longer, so this reading is less certain than the others.
Sources: METR [1]; METR on X [4]. Chart by Kelly Dennis.
The numbers behind this chart
| Item | Value | Note |
|---|---|---|
| Claude 3.7 Sonnet | 1h | |
| GPT-5 | 3.6h | |
| Claude Opus 4.5 | 5.3h | |
| Mythos (early) | 16h | 95% confidence interval: 8.5 to 55 hours |
METR@METR_EvalsWe evaluated an early version of Claude Mythos Preview for risk assessment during a limited window in March 2026. We estimated a 50%-time-horizon of at least 16hrs (95% CI 8.5hrs to 55hrs) on our task suite, at the upper end of what we can measure without new tasks. pic.twitter.com/yIG1Ux27Ro
What 'Half the Time' Actually Means
A fifty percent time horizon is a probability. METR defines it as a task length, timed by a human expert, that a model finishes half the time it tries.1 A model with a two-hour horizon might ace that task once and miss the next attempt. The number describes odds alone.
METR also tracks an eighty percent horizon.1 That is the task length a model clears four tries out of five. Most jobs worth automating need better odds than a coin flip. Reliability decides what a product leader can hand off, far more than raw time does.
Messy Tasks Cut the Clock to Minutes
A different test told a harder story.3 METR gave Claude 4.5 Sonnet an everyday job: make a cup of coffee. The model's horizon for that job measured about two minutes. Clean software tasks on METR's main chart score most frontier models in hours.
The gap gets worse by domain.3 Time horizons run similar for math tasks. They land 40 to 100 times lower for visual, screen-based work, mostly from poor perception. Messy, real-world jobs break agents the same way the coffee test did.
Visual AI Tasks Score 40 to 100 Times Below Math on METR's Chart
How many times lower a visual, screen-based task's horizon runs than a math task's, METR's follow-up study
The blue dot is the low end of the measured range and the magenta dot is the high end; both show how far visual, screen-based tasks fall behind math tasks of the same difficulty.
Source: METR [3]. Chart by Kelly Dennis.
The numbers behind this chart
| Item | Low estimate | High estimate |
|---|---|---|
| Visual, screen-based tasks | 40x lower | 100x lower |
Chasing money produced the same result.2 Anthropic's Opus 4.7 tried to earn cash on its own. It ran four separate times over about four days. Each run ended with an empty account, even after it spotted obstacles like CAPTCHAs.
Inioluwa Deborah Raji is a PhD student at UC Berkeley who studies model evaluation.5 She doubts that task length alone proves anything about difficulty. "I don't think it's necessarily a given fact that because something takes longer, it's going to be a harder task," Raji said.
Her point lands hardest on the coffee test. A task that takes a person two minutes still tripped up a frontier model built mostly on code.
Both Sides on What the Clock Proves
Peter Zaffino, AIG's chairman and chief executive, argues the record already answers the question.6 Speaking on the insurer's first-quarter earnings call, he described the change in plain terms. "When we began our work with Claude 2.0, AI agents could operate autonomously for less than an hour. Today, they can run autonomously for as long as 30 hours," Zaffino said.
AIG built that stretch into real underwriting work, in partnership with Palantir and Anthropic.
Zaffino called the change itself a surprise.7 "While we expected the technology would develop meaningfully over time, we could not have predicted the rapid pace of advancement over the last nine months, or the breadth of AI's potential application," he told the same call.
Daniel Kang teaches computer science at the University of Illinois Urbana-Champaign.5 He reads the same chart and reaches the opposite verdict. METR's suite leans almost entirely on coding and research tasks.
Large language models already train heavily on that kind of work.5 "A model can get better at coding, but it's not going to magically get better at anything else," Kang said.
His caution matters most for the jobs AIG still keeps with human underwriters. Those include claims adjusting, customer calls, and other work that leans on judgment over a verifiable answer.
METR itself sits between the two readings. Mythos's score sat "at the upper end of what we can measure without new tasks," the group wrote when it posted the result.4 Evaluation methods are growing more slowly than the models they test, The Decoder's Matthias Bastian wrote after the score posted.8
What to Do
Match the task to the horizon METR actually measured, past the headline number alone. Hand an agent a coding or research job inside its fifty percent range. Check its work again before you trust the eighty percent range. Save the sixteen-hour claims for a closely watched pilot, before any production handoff.
Listen for the word messy in any vendor's pitch. A model that aces a scored benchmark can still stall on a messy judgment call. That is the same gap METR's own coffee check exposed. Ask for a demo on your actual workflow before you trust a chart built on someone else's tasks.
By the Numbers
- 88.6 days: an agent's clean-task horizon now doubles this fast, down from 196.5 days since 20191.
- 16 hours: the fifty percent success horizon METR measured for an early Claude Mythos Preview, its highest reading ever posted4.
- 8.5 to 55 hours: the confidence range around that score, because only 5 of METR's 228 tasks run that long4.
- 2 minutes: Claude 4.5 Sonnet's horizon on an everyday coffee-making task, from METR's own check on the limits of its benchmark3.
- 30 hours: how long AIG says its agents now run unsupervised in underwriting work, up from under an hour with Claude 2.06.
- 40 to 100 times: how much lower agents score on visual, screen-based tasks than on math, mostly from poor perception3.
- An empty account: what Claude Opus 4.7 earned across four separate autonomous money-making runs over about four days2.
What to Watch
METR says it is building a longer task suite for real, messy jobs past code.3 That update should post a messy-task score next to today's clean one. AIG's quarterly filings are worth a second look too, since its agentic underwriting numbers now run as a named line in its investor materials. The next earnings call will show whether thirty hours of autonomy also means thirty hours of accuracy.
Sources
- METR, "Time Horizon 1.1," METR, Jan. 29, 2026, https://metr.org/blog/2026-01-29-time-horizon-1-1/
- METR, "Frontier Risk Report (February to March 2026)," METR, May 19, 2026, https://metr.org/blog/2026-05-19-frontier-risk-report/
- METR, "Clarifying limitations of time horizon," METR, Jan. 22, 2026, https://metr.org/notes/2026-01-22-time-horizon-limitations/
- METR (@METR_Evals), post on X, May 8, 2026, https://x.com/METR_Evals/status/2052896621760004602
- Grace Huckins, "This is the most misunderstood graph in AI," MIT Technology Review, Feb. 5, 2026, https://www.technologyreview.com/2026/02/05/1132254/this-is-the-most-misunderstood-graph-in-ai/
- AIG, Peter Zaffino, "AIG (AIG) Q1 2026 Earnings Call Transcript," The Motley Fool, May 1, 2026, https://www.fool.com/earnings/call-transcripts/2026/05/01/aig-aig-q1-2026-earnings-call-transcript/
- Beth Musselwhite, "AI advancing faster than expected as AIG builds multi-agentic solution: CEO Zaffino," Reinsurance News, May 1, 2026, https://www.reinsurancene.ws/ai-advancing-faster-than-expected-as-aig-builds-multi-agentic-solution-ceo-zaffino/
- Matthias Bastian, "METR says it can barely measure Claude Mythos, Palo Alto Networks warns of autonomous AI attackers," The Decoder, May 10, 2026, https://the-decoder.com/metr-says-it-can-barely-measure-claude-mythos-palo-alto-networks-warns-of-autonomous-ai-attackers/