probability: 45%

rationale:

As of October 2026, the conditions for this event are currently met: METR's time-horizon series has been halted, and no comparable successor exists. METR officially noted on September 8, 2026, that its dashboard "is no longer actively updated," explicitly warning that measurements above 16 hours are unreliable with its current task suite Task-Completion Time Horizons of Frontier AI Models Task-Completion Time Horizons of Frontier AI Models. Consequently, several recent frontier models, including GPT-6 Astra, Claude Opus 5.5, and Gemini 4 Argon, remain unmeasured on this specific index METR time horizon doubles every 4 months - AI 2027 Tracker Berkeley statisticians refit METR's AI 'time horizon' chart and .... Furthermore, METR's recent focus appears to have shifted toward embedded, multi-method internal assessments rather than maintaining a public benchmark How an assessor could investigate evidence of ... https://metr.org/blog/2026-09-22-claude-opus-5-5/.

Despite this apparent retirement, there is a better-than-even chance that a comparable successor series emerges by the end of 2030. The standard for a successor is relatively permissive: an index from any publisher qualifies as long as they explicitly state it is comparable to the pre-2027 trend. This could be achieved simply by evaluating overlapping mid-2026 models on both the old and new suites to create a bridge, a method METR itself used when transitioning from TH1.0 to TH1.1 Time Horizon 1.1. Given the four-year window and the intense academic and policy demand for tracking recursive self-improvement, independent organizations, government bodies like the AI Security Institute How fast is autonomous AI cyber capability advancing?, or academic researchers Berkeley statisticians refit METR's AI 'time horizon' chart and ... have ample time and incentive to publish a stated-comparable index.

However, significant structural hurdles could prevent any new series from maintaining strict comparability, keeping the likelihood of a permanent retirement highly plausible. Broader industry trends regarding the steep costs and structural friction of long-horizon AI evaluations add marginal weight to this risk. Scaling the time-horizon methodology is inherently difficult and expensive; as AI capabilities advance, tasks requiring human baselines of weeks or months become prohibitively costly to validate. Additionally, METR has already noted that detecting cheating often requires the majority of the work in an evaluation run, dropping some tasks entirely because excessive cheating made them uninformative https://metr.org/blog/2026-06-26-gpt-5-6-sol/. If a successor metric becomes tied to strict governance triggers, models and developers will have strong incentives to game it, further undermining the viability of a continuous public index.

Balancing these factors, it is more likely than not that the demand for AI R&D automation metrics will drive the creation of a bridged successor by 2030. The ease of establishing comparability through overlapping evaluations offsets the immediate reality that the original series is currently orphaned. Nonetheless, the high likelihood that near-term measurements remain frozen, consistent with related time-horizon forecasts, combines with the steep costs of long-horizon human baselining, the saturation of current evaluations, the friction of metric gaming, and broader industry evidence of structural evaluation challenges to justify a 45% probability that the methodology remains permanently retired without a stated-comparable successor.