Anthropic shipped one model with two doors, and only one of them opens for you
Hi there,
A thousand AI agents built a working operating system from scratch in eight days. That is the story Anthropic told on Monday, when it released Claude Fable 5.2. The agents worked together, wrote the code, fixed their own mistakes, and ended up with something that runs. Anthropic says it is one of the tests its early users ran. Nobody outside the company has checked it yet, so take it as their story, not a fact. Still, it is a strange thing to picture: a thousand copies of one model, building a whole system while nobody watches.
The model itself is Anthropic's most capable one for hard technical work. It writes code, reasons through long problems, and works with agents. It costs $8 for a million input tokens and $40 for a million output tokens, down from $10 and $50, and Anthropic says it answers up to 30 percent faster than the last version. It also holds a million tokens of context, which is a lot of room to think in.
Here is the part that stayed with me. Fable 5.2 is one model with two doors. The one called Fable is open to everyone, with safety blocks that stop most cybersecurity and biology work. The one called Mythos is the same model with those blocks lifted, and it only opens for organisations Anthropic has checked and approved. Same brain, different keys. Anthropic is saying, in plain terms, that the strongest version is not for everyone, and it is not pretending otherwise.
There is a small honest note in the announcement too. Anthropic still tells most people to start with Claude Opus 5.5 for everyday coding and knowledge work, and to reach for Fable only for the biggest, most open-ended jobs. A company telling you to buy the cheaper thing first is not something you hear every day.
Pro users get to try Fable 5.2 at no extra cost until November 16, though it can use up to half of their weekly allowance. It also lands in the week Anthropic is reported to be close to a stock market listing, which explains the timing, if not the model.
If I tried it first, I would give it one long, messy coding job, the kind with many files and no clear starting point, and see whether it holds the whole thing in its head or loses the thread halfway. That is the test that matters to me. What would you give it first?
Cover image: AI-generated