x.com/Plinz/status/2064528835329286487

Anthropic’s own June 9 announcement contradicts this in two ways.

1. The generally available model was not “public Mythos.” Anthropic said it was launching Claude Fable 5 for general use, while Claude Mythos 5 would be available only to a small group of cyberdefenders and infrastructure providers through Project Glasswing. So taken literally, there was no general-public release of “Mythos” itself.

2. Anthropic says the public release is stronger, not weaker, than prior public models. In the same announcement, Anthropic says Fable 5’s capabilities “exceed those of any model we’ve ever made generally available,” that it is “state-of-the-art on nearly all tested benchmarks,” and that it shows “exceptional performance in software engineering.” Those statements directly conflict with the idea that the public release was designed to perform worse than its predecessors on technical tasks in general.

3. Anthropic had already described an earlier public model as the first one with deliberately reduced cyber capabilities. In the April 16 announcement for Claude Opus 4.7, Anthropic wrote that Opus 4.7 was “the first such model” and said that during training it had “experimented with efforts to differentially reduce” its cyber capabilities. That means even if the post is referring specifically to deliberately reduced cyber-related capabilities, Anthropic’s own description says the first such public model was Opus 4.7, not the June 9 Mythos-class release.

What is true is that Anthropic added safeguards to Fable 5: some cyber/bio/distillation queries are routed to Opus 4.8 instead. But that is different from saying the model is the first one deliberately designed to perform worse than its predecessors on technical tasks overall.