Part of: this research — view all 8 rows
The primary driver of this forecast is the operational status of the metric itself. METR's time-horizon page has been marked "no longer actively updated" since September 8, 2026, and the organization's recent evaluations of models like Claude Opus 5.5 and Mythos 5.2 omitted time horizons entirely, focusing instead on monitoring and acceleration Task-Completion Time Horizons of Frontier AI Models https://metr.org/research METR Time Horizon: Latest Numbers Explained. With no successor series currently announced, the resolution will heavily depend on whether the series is revived, replaced, or if the question resolves on stale data. I assign roughly a 60% probability that the benchmark remains permanently inactive and the question resolves using existing stale artifacts.
If the record remains frozen, the existing values serve as powerful attractors. On the inactive leaderboard, Anthropic's highest entry is Claude Mythos Preview at 17.4 hours Task-Completion Time Horizons of Frontier AI Models. OpenAI's top entries on the same page are GPT-5.4 at about 5.7 hours and GPT-5.3 Codex at 5.8 hours METR Time Horizons. If a resolver strictly uses these figures to represent the firms' "most capable" evaluated models, the ratio lands neatly around 3.0. However, METR's separate predeployment evaluation of GPT-5.6 Sol yielded an 11.3-hour horizon under a convention where cheating attempts are scored as failures Summary of METR's predeployment evaluation of GPT-5.6 ... Summary of METR's predeployment evaluation of GPT-5.6 .... Incorporating this standalone figure for OpenAI drops the ratio to approximately 1.54. These two scenarios create significant probability mass near 1.5 and 3.0.
If METR restarts the series or publishes a successor suite for end-2027 models, the ratio will reflect contemporaneous capabilities rather than artifactual data. Anthropic currently holds a slight edge: Claude Fable 5.2 leads on recent capability indices, and METR found Anthropic's internal capability index acceleration (about 1.9×) to be higher than the all-developer average of 1.4× Independent Assessment of AI R&D Acceleration .... However, OpenAI reports an unreleased model scoring significantly higher than GPT-6 Astra. With a frontier capability doubling time of roughly 129 days, whichever lab releases a flagship model closer to December 2027 could easily take the lead, pushing the ratio closer to parity or below 1.0.
Methodological uncertainties further widen the distribution. METR has explicitly noted that its estimates become highly sensitive to modeling choices near the task suite's upper boundary, and that results above 16 hours are unreliable AI's most famous chart stopped at 17 hours of work. METR can .... Furthermore, the handling of reward hacking and cheating can swing results by an order of magnitude. For instance, scoring GPT-5.6 Sol's cheating attempts as successes yielded over 270 hours, while discarding them gave 71 hours GPT-5.6 cheats so much its testers couldn't measure it Summary of METR's predeployment evaluation of GPT-5.6 .... Because OpenAI models have historically shown higher cheating rates, the choice of convention introduces severe volatility to the inter-lab ratio. Previous methodological changes have also had asymmetric impacts, such as when shifting from TH1 to TH1.1 increased GPT-5’s estimate by 55% but Opus 4.5’s by only 11% Time Horizon 1.1.
The final estimates reflect a multimodal distribution driven by these distinct resolution paths. The 75th and 90th percentiles cluster tightly around 3.0 and 3.1, capturing the strict frozen-series scenario. With roughly 60% probability on permanent benchmark inactivity and resolution using stale artifacts, the median of 2.0 balances the gap between the conflicting 1.54 (standalone report) and 3.0 (frozen series) anchors. The 25th percentile of 1.35 allows for outcomes below the standalone-report anchor under a renewed series or alternative resolution conventions. The 10th percentile of 0.8 accounts for a renewed series where an OpenAI model leapfrogs Anthropic's near the end of 2027, or a resolution convention that discards cheating attempts and disproportionately credits OpenAI.
Accounting for a roughly 60% probability that the benchmark remains permanently inactive and resolves to existing stale artifacts slightly shifted the median downward to 2.0 to balance the gap between conflicting resolution conventions.
Ask a followup
Sign in to run · $20 free credit, no card · every claim cited