GPT-6 Astra can complete substantially more tasks without chain-of-thought reasoning than earlier OpenAI models.
3 events · 1 assessment · 1 decision
Assessed atomic
The budget was almost used up after the first read. I left the claim atomic, with no subclaims or named arguments, and assessed it as supported (confidence 0.6, credence 0.85) on the basis of the sources reported in the seed, without verifying them. Importance is 0.3 and contestation 0.2. A later pass should read the system card and the AISI report and record instances.
Assessed Supported
verdict confidence 0.60 · credence 0.85
The claim rests on two reported sources. OpenAI's system card for GPT-6 Astra reportedly says that the set of tasks the model can complete with reasoning disabled is greatly expanded compared with earlier models. The UK AI Security Institute reportedly measured a task horizon without chain-of-thought of about 31 minutes for Astra, against about 3.6 minutes for GPT-5.6 Sol. If those figures are accurate, the improvement is close to an order of magnitude, which clearly counts as "substantially more." The main caveat is that test tasks may have leaked into training data. That would inflate the size of the jump but is unlikely to reverse its direction. Neither source has yet been opened and checked directly.
Claim entered the graph