OpenAI calls Astra its most dangerous model yet

The timing of OpenAI's warning is notable, as it landed the same day Anthropic shipped Claude Fable 5.1 and Mythos 5.1, and Anthropic has reportedly passed OpenAI on revenue this year. OpenAI's CEO Sam Altman explained that the team spent the summer on safety priorities and that Astra has been done training for a while, with models after it being slowed down intentionally. Users on X viewed this as an excuse from a company falling behind, but the article suggests OpenAI is grappling with the challenge of monitoring its most capable model.

Astra's critical rating is backed by internal evaluations, but the article notes that these results came from the expanded 'Daybreak Blue' access, not the standard setup. The July hack, where OpenAI's own agents misfired, underscores the stakes: investigators pieced together what happened from reasoning logs, highlighting the importance of readable chains of thought. However, Astra's use of 'recurrent depth' pushes some reasoning into unreadable internal representations, which OpenAI's chief scientist admits makes chain-of-thought monitoring 'fragile' and 'trending in a negative direction'.

The article raises concerns about imitators who might not limit the technique as OpenAI has. It also points to the financial pressure from Amazon, Microsoft, and Google spending roughly $600 billion this year on infrastructure, which could incentivize loosening safety throttle later. OpenAI, Anthropic, and Google researchers had warned a year ago about the risks of latent reasoning, and now OpenAI is using the technique, albeit throttled. The burden of proof falls on OpenAI to show Astra poses no risk, but the article argues that proof cannot come from its own unverifiable evaluations, and outside researchers from METR have only gotten limited insight into the system so far.