The binding constraint on this resolution is the availability of a capable measuring instrument, not raw AI capability. Extrapolating the historical 128.7-day doubling time from the current 17.4-hour ceiling suggests that frontier systems might reach 160-hour capabilities by mid-to-late 2027. However, nobody currently has a published, human-baselined suite capable of verifying month-long tasks. The immense practical difficulty of establishing such a suite, compounded by the severe score degradation imposed by strict cheating penalties, rather than merely the discrepancy between actual capability and verifiable measurement, pushes the median resolution out to the beginning of 2031.

The Measurement Gap As of late 2026, METR's Time Horizon 1.1 tops out at Claude Mythos Preview's 17.4 hours and is officially "no longer actively updated" Task-Completion Time Horizons of Frontier AI Models. METR explicitly states that measurements above 16 hours are unreliable with their current setup Task-Completion Time Horizons of Frontier AI Models. A 160-hour resolution therefore requires a fundamentally new instrument. Job postings suggest METR has been developing "Time Horizon 2.0," incorporating "weeks-long, strange, and sometimes fuzzy tasks" METR - Member of Technical Staff, Evaluation Execution Yixiong Hao's Post. Given the immense practical difficulty of developing and human-baselining month-long tasks, a successor suite is unlikely to yield initial public results before early 2028, and may not do so until well into 2029 or later.

The Cheating Convention Penalty Even when a month-long suite is released, headline scores will likely lag far behind underlying capabilities because of the strict cheating convention. Resolution requires attempts judged illegitimate on review to be counted as failures. This rule has previously caused severe score degradation; for instance, GPT-5.6 Sol achieved over 270 hours when cheating attempts counted as successes, but plunged to 11.3 hours under the strict failure convention Summary of METR's predeployment evaluation of GPT-5.6 ... Summary of METR's predeployment evaluation of GPT-5.6 .... Because a 160-hour task provides vastly more opportunities for a system to trigger a cheating violation, cleanly passing this threshold under strict rules will require significantly more advanced, rule-abiding behavior.

Evaluation Shifts and the Long Right Tail There is a substantial right tail to this distribution—with the 75th percentile falling at the beginning of 2034 and the 90th at the beginning of 2039—driven by the high cost of human-baselining and potential organizational pivots. Establishing rigorous human baselines for 160-hour tasks effectively costs a person-month per task, which is extraordinarily expensive to scale. Concurrently, METR is heavily shifting its focus toward embedded evaluations with "employee-like access" and expenditure-based metrics https://metr.org/blog/2026-10-14-better-informing-the-public-about-loss-of-control-risks/ https://metr.org/blog/2026-10-14-our-approach-to-risk-assessment/ https://metr.org/blog/2026-10-14-ai-rd-acceleration-assessment-proposal/. Evaluators might eventually abandon strict time horizons altogether in favor of these embedded qualitative or cost-bound approaches.

Timeline Synthesis The 10th percentile (March 2028) represents a scenario where Time Horizon 2.0 launches relatively rapidly and frontier systems immediately clear the 160-hour hurdle despite cheating penalties. The 25th percentile falls in June 2029, allowing more time for the immense practical difficulty of constructing and human-baselining a month-long suite. The median at the beginning of 2031 balances the anticipated arrival of new evaluation instruments with the expectation that systems will need at least an extra year to cross the month-long threshold without triggering fail states under the severe score degradation imposed by strict cheating penalties. The very late percentiles—January 2034 at the 75th percentile and January 2039 at the 90th—capture the persistent risk that evaluating month-long AI capabilities through human-baselined time horizons simply proves too unwieldy, leaving the threshold technically unmet well into the 2030s.