This market will resolve to yes if any model is released in 2026 that has a METR time horizon (50% reliability) of at least 8 hours. It will resolve to no, if at the end of 2026, no such model has been released.
1,000
1.00https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/
So a "doubling time" of about 260 days (8.7 months) would be required to reach 8 hours in 2026 if we extrapolate from GPT-5.1 Codex Max (2h42m), or 280 days (9.4 months) from GPT-5, if my maths is correct.
@Jw we've had a doubling time faster than that for a few years now. I default to thinking it's pretty likely that continues.
The other dynamic here favoring a "no" resolution is that the benchmark is only designed to measure tasks up to 8 hours so maybe that makes getting to 8 hours unusually difficult without an update to the benchmark.