What will the highest score in METR's published time-horizon series be on 2029-12-31, in hours at the 50% success rate, for a model from OpenAI, Anthropic or Google DeepMind?

rationale:

The Median Case: A Frozen Series

The median estimate of 17.42 hours is firmly anchored to the status quo. Evaluated alongside related questions on this methodology, it reflects a roughly 55% chance that the series remains frozen through 2029. As of September 2026, METR explicitly marked its time-horizons page as "no longer actively updated" Task-Completion Time Horizons of Frontier AI Models Task-Completion Time Horizons of Frontier AI Models. The highest current entry is Anthropic's Claude Mythos Preview (early) at 17.4 hours, added in May 2026. METR's Frontier Risk Report from that month indicated that the strongest agents had essentially saturated the Time Horizon 1.1 tasks https://metr.org/blog/2026-05-19-frontier-risk-report/ https://metr.org/blog/2026-05-19-frontier-risk-report/. Furthermore, METR has stated that "measurements above 16 hrs are unreliable with our current task suite" Task-Completion Time Horizons of Frontier AI Models Task-Completion Time Horizons of Frontier AI Models. Because organizations rarely un-archive deprecated pages without a complete methodology overhaul, there is a better-than-even chance (around 55%) that the published series never receives an eligible update, leaving the high score effectively frozen at 17.4.

Shift in Assessment Methodology

METR's recent actions strongly suggest they have moved away from this specific metric for new frontier AI models. Major releases from late 2026—including OpenAI's GPT-6 Astra, Anthropic's Claude Opus 5.5, and Google's Gemini 4 Argon—do not appear on the leaderboard METR time horizon doubles every 4 months - AI 2027 Tracker METR Time Horizons. When evaluating Claude Opus 5.5 in September 2026, METR provided no time-horizon figure, opting instead for five specific AI R&D tasks and a holistic assessment Summary of METR's predeployment evaluation of Claude ... https://metr.org/blog/2026-09-22-claude-opus-5-5/. Similarly, a predeployment evaluation of GPT-5.6 Sol yielded 11.3 hours (or 71 hours if detected cheating runs were discarded), but METR explicitly cautioned that neither was a robust measurement Summary of METR's predeployment evaluation of GPT-5.6 ... Summary of METR's predeployment evaluation of GPT-5.6 .... With their recently announced $71M funding, METR appears to be pivoting toward embedded, interview-based assessments of loss-of-control risk rather than extending the v1.1 task suite A proposal for better informing the public about loss-of- ... https://metr.org/blog/2026-08-14-funding-update/.

The Right Tail: A Successor Series

The heavy right tail—placing the 75th percentile at 65 hours and the 90th percentile at 500 hours—accounts for a 35–40% chance that METR eventually launches a formal "Time Horizon 2" or equivalent successor series that counts toward resolution. AI capabilities have been growing rapidly; METR previously reported a post-2023 doubling time of roughly 129 days Task-Completion Time Horizons of Frontier AI Models. Broader expectations for further AI capability leaps and R&D automation support somewhat longer time horizons for a successor metric if one is introduced. If we purely extrapolate a 4–7 month doubling time from April 2026 out to late 2029, the implied time horizons reach into the thousands of hours. However, in practice, building and scoring tasks that take humans weeks or months to complete is extremely expensive and difficult. Measurement constraints, rather than pure capability, will likely cap reported values. Therefore, if a new series is published, scores will realistically land in the tens to hundreds of hours.

Downward Revisions

The lower percentiles sit at or just below the 17.4-hour mark, at 17.2 hours for the 10th percentile and 17.4 hours for the 25th percentile, reflecting a small probability (around 5%) of downward revisions. METR has previously adjusted its scores to correct statistical errors that had artificially inflated some 50% horizons. If they issue a minor correction to the current page or release a successor suite that re-baselines task difficulty downward—noting that their past v1.0 and v1.1 values were not directly comparable—the highest standing score on the established series could dip slightly below its current peak.