METR AI time horizon scores have been doubling every 7 months for a while now. I will give slightly more leeway by only requiring an average of 8 months per doubling over the course of 3 doublings. A date will resolve yes if no AI model exists with an 80% time horizon greater than 8x the best released model at that date two years after that date.

For example, Claude Opus 4.6 has a METR 80% time horizon of 63 minutes and was released in Feb 2026. If by February 2028 there is no model with an 80% METR time horizon of more than 63*8 = 504 minutes, this market will resolve to 2026.

If METR doesn't exist anymore, I will try to find an equivalently trustworthy source (not one sponsored/owned by the companies whose models they test) with a similar measurement.