TAWS

From the Tuesday, November 10 show

Claude Fable 5.2: up to 30% faster, with a deliberate cap on AI-building skills

AI-generated from the transcript

What happened
Anthropic released Claude Fable 5.2 for paid Claude plans, with businesses paying per use paying 20% less. Outside testers including the UK AI Security Institute and METR checked the model before release. In September, CEO Dario Amodei called for slowing the pace of AI development.
The claim
Anthropic says it's up to 30% faster than the last version. It says it held back skills like tuning code for AI chips and improving AI model designs, keeping them around Opus 5.5 level, while ordinary coding and traditional AI research weren't limited. Dianne Penn told CNBC 'we called for pacing, not pausing.' Anthropic says Claude writes over 90% of changes merged into its core code base, with human guidance. Its safety report puts AI progress speedup at about 1.88x against a 2x line, and says Mythos doesn't seem close to replacing research staff.
Austin's take
Verify. Enjoy the faster answers, but make Anthropic earn trust on the speed limit. About a third of the safety reruns hit 2x anyway, and the strongest case for a lasting speedup leans on an unreleased model you can't use. On the staff survey: sometimes Claude brings a lab coat to a napkin problem. Ron argued independent testing plus a public self-own is what trust looks like.
Why it matters
AI helping build better AI is the risk safety rules are designed around, and this is a company grading how close it is to its own line.
What to do
Paid Claude users can try the faster model now. Look for more outside, independent checks on the capped skills before taking the cap at face value.
Not confirmed yet
The speedup figures come from Anthropic's own safety report, and some reruns crossed the 2x line; the report also scores a stronger unreleased model.

Watch from 12:55 on the episode page

The Loop

The weekly newsletter

The week in AI in one email: what happened, why it matters, what to do about it.