Anthropic

On 1 September 2026, Anthropic shipped Claude Fable 5.1 and Claude Mythos 5.1 as one underlying model at two guardrail levels: Fable 5.1 in general availability, Mythos 5.1 reachable only through trusted-access programmes, and still below the next Responsible Scaling Policy risk threshold in biology.

That split is the company's signature move. Anthropic lodges the guardrail inside the weights and arbitrates who is allowed past it, the inverse topology of a separable, self-hostable classifier. Read unsympathetically, the arrangement is a permission regime selling cognition as infrastructure.

The other reason the notes keep circling back is that Anthropic is its own most documented deployment. Jason Clinton, Deputy CISO, reports that Claude writes about 80% of merged code, and that substantial review comments rose from 16% to 54% of pull requests once the Test stage was rebuilt as the identified bottleneck. Louis Claxton's The AI-Native SDLC playbook generalises the same loop into committed artefacts. Boris Cherny and Cat Wu describe product and engineering roles merging inside the team. Every one of those numbers comes from the vendor of the model, measured on a young codebase.

The political position is unsettled. Anthropic refused to let Claude serve autonomous warfare and mass surveillance, saw federal use of its technology barred by presidential order, sued, and lost an attempt to have the blacklisting blocked on appeal. On the commercial side it announced a two-pool split of Agent SDK billing, then suspended the change the day it took effect: "For now, nothing has changed."