The longest task horizon METR has measured for any AI model is about 17.4 hours, for Claude Mythos Preview (early), according to METR’s own data file. The updated METR graph spreading on X since 16 November adds points far above that, led by an Anthropic internal model at about 240 hours, which is an estimate rather than a METR measurement.
That 240-hour figure comes from Nikola Jurkovic, whom METR’s team page lists as former technical staff. Jurkovic derived it on 6 November from capability scores in METR’s assessment of Anthropic and called it “pretty weak evidence”. METR’s time-horizon page has not been updated since 8 September and warns that measurements above 16 hours are unreliable.
Business Pill · AI Agents
An evergreen Business Engineer explainer on AI agents, from answering to doing. It teaches the general idea only and does not discuss METR, time horizons or Anthropic’s models.
The key insight: As we read the sources, the viral chart mixes two kinds of points: METR measurements that stop at about 17.4 hours, and capability-score conversions for Anthropic internal models that METR did not run through its task suite.
What METR Has Actually Measured
METR defines the time horizon as the task length, measured by how long human experts take, at which an AI agent is predicted to succeed with a given reliability. Its raw data file for Time Horizon 1.1 gives Claude Mythos Preview (early), dated 7 April 2026, a 50% horizon of 1,044.8 minutes and an 80% horizon of 185.9 minutes.
By our arithmetic that is about 17.4 hours at 50% success and 3.1 hours at 80%. The 50% estimate has a wide confidence interval in the file, from 508.9 to 3,304.3 minutes. The next-highest record in the file is Claude Opus 4.6 at 718.8 minutes, or about 12 hours.
METR added Mythos Preview (early) on 8 May with the notice “Measurements above 16 hrs are unreliable with our current task suite”. On 8 September it “Clarified that this page is no longer actively updated”. So no model after Mythos Preview has a METR time-horizon measurement on that page.

Where the 240-Hour Point Comes From
On 5 November METR published an assessment of AI R&D acceleration at Anthropic, which we covered in METR’s finding that Anthropic is likely improving models at about twice its old rate. It tracks Anthropic models with AECI, “Anthropic’s version of the Epoch Capabilities Index”, including internal models up to Model 4, which METR says finished training in September to October 2026.
The assessment gives no time horizon in hours for any internal model. Jurkovic posted a chart a day later converting AECI scores into 50% horizons, writing that they “imply that the time horizon doubling time is now around 2 months, and that the TH of Model 4 is around 240 hours”. His chart fits a 56-day doubling from Opus 4.5 onward and labels its AECI-derived points as estimates.
He added: “It’s unclear if there is a clear 1:1 mapping from AECI to time horizon for numbers this large”. By our arithmetic, 240 hours is about 13.8 times the highest horizon METR has published, and 15 times the 16-hour level METR calls unreliable.
How the Viral Chart Turns 240 Hours Into 38
Benjamin Todd’s 16 November post, “The METR graph, updated:”, plots 80% horizons. Its legend marks the internal-model diamonds as “Not measured” and says they were estimated from AECI as the 50% horizon divided by 6.3, which is how Model 4 appears at about 38 hours. The post had about 74,000 views when we read it.
METR’s own figures give a different ratio for the one model near that range: Mythos Preview (early) has a 50% horizon about 5.6 times its 80% horizon, by our arithmetic. Todd’s chart also draws a trend of doubling every 1.7 months since Opus 4.5 that includes the estimates, next to METR’s long-run 7-month line.
The Gap in One Number
METR says its original time-horizon work found a 7-month doubling time. At that pace a horizon grows about 3.3 times in 12 months, which would take Mythos Preview’s 17.4 hours to about 57 hours, by our arithmetic. At the 56-day doubling in Jurkovic’s chart, 12 months is about a 92-fold increase.
Going from 17.4 to 240 hours takes about 3.8 doublings, or roughly 26 to 27 months at METR’s 7-month rate, by our arithmetic. Third-party evaluation access has been in the news this month, from OpenAI opening its models to four safety labs to whether a US sign-off is needed before a launch.
The Structural Read
As we read METR’s own pages, the measured record has stood since Mythos Preview (early) was added on 8 May, and METR flags anything above 16 hours as unreliable on its current tasks.
As we read Jurkovic’s post, the 240-hour figure carries its own caveat: he wrote that it is unclear whether AECI maps one-to-one onto time horizons at numbers that large.
Nikola Jurkovic on X, 6 November 2026
“TBC this is pretty weak evidence”
Three Implications
FOR READERS OF THE CHART Benjamin Todd’s legend marks the diamond points as not measured and converts them from 50% to 80% horizons by dividing by 6.3.
FOR THE TREND LINE METR’s original work found a 7-month doubling; at that rate a horizon grows about 3.3 times in a year, by our arithmetic, against about 92 times at the 56-day fit in Jurkovic’s chart.
FOR ANTHROPIC’S PACE METR’s 5 November assessment estimates Anthropic’s AECI scores have accelerated 1.9 times since October 2025, with a 95% confidence interval of 1.5 to 2.8 times.
The Business Engineer Lens
This story maps onto the Business Engineer framework The AI Supercycle.
The framework describes the “Short cycle (5–10 years): rapid capability improvements and market disruptions. Where the bubble lives.”
As we read it, a time-horizon chart is a short-cycle instrument, and the gap between a measured 17.4 hours and an estimated 240 hours is the kind of reading the framework keeps separate from the longer cycles. The framework does not discuss METR, Anthropic or time horizons.

What Is Not Established
No one has measured Model 4’s time horizon on METR’s task suite, as far as METR’s public pages show. The 240-hour and 38-hour figures are conversions of capability scores read from published charts, and Jurkovic himself questioned the mapping at that scale. The ratio of 6.3 used for the 80% chart is Todd’s; we did not find its derivation.
METR’s 5 November assessment says Anthropic’s internal benchmarks are increasingly saturated. Neither METR, Anthropic, Jurkovic nor Todd was contacted for this piece.
The Bottom Line
METR’s highest published time horizon is 17.4 hours at 50% success, for Mythos Preview (early), and its page has been frozen since 8 September. The 240-hour Model 4 point in the viral graph is a former METR staffer’s AECI-based estimate that he called pretty weak evidence.
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. We read METR’s time-horizons page and its linked Time Horizon 1.1 data file, METR’s 5 November 2026 assessment of AI R&D acceleration at Anthropic, and METR’s team page, all in full. We read Nikola Jurkovic’s 6 November post on X and its chart, and Benjamin Todd’s 16 November post and its chart. Hours, ratios and growth factors are our arithmetic from those figures. The AECI-derived horizons are third-party estimates, not METR measurements. METR, Anthropic, Jurkovic and Todd were not contacted for this piece.
Sources: METR: Task-Completion Time Horizons of Frontier AI Models · METR Time Horizon 1.1 data file · METR: Independent Assessment of AI R&D Acceleration at Anthropic (5 Nov 2026) · METR team page: Nikola Jurkovic · Nikola Jurkovic on X (6 Nov 2026) · Benjamin Todd on X (16 Nov 2026)









