GPT-6 Astra's no-chain-of-thought task horizon exceeds the level predicted by earlier models at its general capability level.
3 events · 1 assessment · 1 decision
Assessed atomic
Budget-constrained first pass (about $0.41 per turn, few turns available). Kept claim atomic with no named arguments; assessed supported (conf 0.5, credence 0.6) from the parent-recorded AISI measurement and informal baseline. Primary sources not re-opened; marginal yield 0.5 so a later pass should read the AISI report and any capability-vs-horizon trend fit.
Assessed Supported
verdict confidence 0.50 · credence 0.60
The claim compares two things: GPT-6 Astra's measured task horizon when it answers without chain-of-thought reasoning, and the horizon earlier models would lead one to expect from a model of Astra's general capability. On the first, the UK AI Security Institute reports about 30.9 minutes for Astra, against 3.6 minutes for its predecessor GPT-5.6 Sol. That is roughly an eightfold jump in one generation. On the second, commentators such as Zvi Mowshowitz estimate that Astra's general capabilities would predict only about 5 to 9 minutes. If both figures hold, the measured horizon is three to six times the expected level, so the gap is large. Two things limit confidence. First, the expected level is an informal extrapolation from only a few earlier models, not a fitted trend with stated uncertainty. Second, the institute itself flags possible contamination of its task set and limited testing time, either of which could inflate Astra's figure. A formal fit of no-chain-of-thought horizon against a general capability index across many models, plus a replication on uncontaminated tasks, would settle the question.
Claim entered the graph