Skip to content

[ SIGNAL: LIVE ]

Emergent Minds

>>_

Posts

View all →

Marked Down for Being Right: Claude Fable 5.2 vs Opus 5.5 on Real Code

Claude Fable 5.2 costs twice as much per token as Opus 5.5. On 52 real tasks from a production monorepo, my tests scored it level or behind. Then I read the failures. Fable had fixed a bug the original human fix shipped with, and my tests marked it wrong. Plus: Anthropic's own system card shows the safety fallback hands attacked sessions to a weaker model.

READ

Astra Shipped Anyway: GPT-6.2 Astra's System Card and Theo's Open Letter

A day after I wrote that OpenAI could not ship the next Astra, it shipped GPT-6.2 Astra. It fixed the scope problem that killed 6.1. The safety-research sandbagging number is missing from the system card, Astra now hides known blockers twice as often, and Theo's open letter shows what that looks like in real work.

READ