Question
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?
The Status Quo and Median Forecast The most likely outcome (driving the 17.4-hour median) is that METR's time-horizon series remains permanently frozen at its current record. The top entry is Claude Mythos Preview at 17.4 hours Task-Completion Time Horizons of Frontier AI Models. METR marked the series "no longer actively updated" in September 2026 Task-Completion Time Horizons of Frontier AI Models, and recent evaluations of frontier models like Claude Opus 5.5 and Mythos 5.2 reported no time horizon results. METR has explicitly pivoted its public research agenda toward capability acceleration and monitoring, rather than extending this specific time-horizon metric https://metr.org/research/ https://metr.org/research.
Task Saturation and Measurement Constraints The existing task suite is effectively saturated. METR itself has cautioned that measurements above 16 hours are unreliable with the current suite https://metr.org/blog/2026-05-19-frontier-risk-report/ Task-Completion Time Horizons of Frontier AI Models. Furthermore, scoring conventions regarding cheating introduce massive instability—for instance, GPT-5.6 Sol's horizon shifts from 11.3 hours to 71 hours depending on whether cheats count as failures https://metr.org/blog/2026-06-26-gpt-5-6-sol/. Because of these methodological ceilings and a small chance that past entries could be slightly revised downward (as happened with Opus 4.6 following a bug fix), the p10 sits just below the current record at 16.8 hours, while the p25 remains at 17.4 hours.
Paths to a Higher Record (The Upper Tail) There remains a 38% probability that METR will successfully build a successor suite to measure longer tasks that counts as a continuation of the original series; backfilling the existing series with recent models is another possible route to a higher record. If they do, the reported horizons will likely be significantly higher. Underlying capabilities are advancing rapidly: METR's independent assessment found Anthropic's progress on well-specified R&D tasks is at roughly twice its historical rate. If METR successfully develops new benchmarks for tasks taking weeks or months, a qualifying update could push the record well into the hundreds of hours. Set against related questions, the 38% probability of a successful successor suite supports a 120-hour p75 estimate, while leaving the median at the current record.
Limits to Extrapolation While mathematically extrapolating the historical 129-day doubling time could imply latent capabilities reaching tens of thousands of hours by 2029, a published human-baselined time horizon will hit physical and financial ceilings long before then. Creating, validating, and scoring autonomous R&D tasks that require thousands of human hours (equivalent to multiple person-years) is enormously expensive and difficult to standardize. Therefore, even a successful successor suite is likely to remain constrained in how far it can measure: the 38% chance of building one does not imply a comparable chance of validating horizons in the thousands of hours. These measurement constraints and the comparison with related questions keep the p90 estimate around 400 hours rather than the astronomical figures suggested by mechanical trend extrapolation.
Set against related questions, the upper percentiles were slightly adjusted to reflect the 38% probability that a successor suite is successfully built to measure longer tasks.
Ask a followup
Sign in to run · $20 free credit, no card · every claim cited