This market will resolve to the highest 50% time horizon, as reported by METR, for any R2 model released within a month of the first R2 announcement.
This market considers that a model released by DeepSeek named V4-thinking would count as R2. V4 would also count if it was natively a reasoning model. If separate models are released, one called V4 and one called R2, R2 will be considered for the purpose of this market. Generally the intent is for this market to resolve according to the model that is the next major iteration after R1.
50% time horizon is a measure of AI autonomy based on the length of tasks that AI can do: roughly, it is the time that humans take to complete tasks that an AI system can successfully do 50% of the time. See METR's "Measuring AI Ability to Complete Long Tasks" for the technical definition. Claude 3.7 Sonnet, released in February 2025, was the leading model with a 50% horizon of 59 minutes.
Left bounds inclusive, right bounds exclusive.
See also:
/jim/claude-45-opuss-metr50-horizon (jim's version)
/Bayesian/claude-opus-45s-metr50-time-horizon (my version)
/Bayesian/gemini-3s-50-time-horizon-per-metr
/Bayesian/grok-420s-metr-50-time-horizon
/Bayesian/claude-sonnet-46s-metr-50-time-hori
/Bayesian/grok-5s-50-time-horizon-per-metr
/Bayesian/r2s-50-time-horizon-per-metr (this market)
1,000People are also trading
The members of the AI futures project have given an update and they appear to now be relying on the 80% time horizon length graph from METR for their predictions rather than the 50% time horizon length graph. This implies that a 50% time horizon is not enough. While I think markets for 50% time horizons are useful, I now think that more attention needs to be paid to 80% time horizon lengths. I am planning to create markets for 80% time horizons as soon as possible unless someone beats me to it.
If that happens, @traders do you agree it's fair to make it about V4 instead? ie if V4 is a reasoning model, R2 would refer to V4-thinking for the purpose of this market?
DeepSeek-R1: 27 mins, released 01-20-25 (SOTA since December was 39 mins)
DeepSeek-R1-0528: 31 mins, released 4 months later (SOTA since April was 1.5 hours)
quadrupling from 31 mins to > 2 hours in another 4 months seems (very) unlikely, not betting more because of uncertainty over when (if ever) it’ll be released.