Claude Fable 5.2 is Anthropic's best model. Its maker says start lower
Anthropic's newest model can design an electronic device from scratch, and its own maker says you probably shouldn't use it for most of your work. That is the most useful line in the announcement of Claude Fable 5.2 and Mythos 5.2, published on 9 November 2026.
Disclosure: SmartHolol uses Anthropic's Claude models, including to help write this blog. Weigh what follows accordingly. Every figure below comes from Anthropic's own announcement, not from independent testing.
The short version
- Fable 5.2 is Anthropic's most capable released model, now faster and cheaper than the version it replaces, but still built for hard, open-ended technical work.
- Anthropic itself recommends starting most coding and knowledge tasks on Opus 5.5, then moving up or down as needed.
- For a business, the lesson is not "use the best model". It is "route each task to the cheapest model that does it well".
What Anthropic announced
Fable 5.2 and Mythos 5.2 are the same underlying model. Fable carries strong cybersecurity and biology safeguards and is available to all users. Mythos is offered to organisations in approved cybersecurity and biomedical research programmes.
Anthropic says the new versions generate output up to 30% faster than 5.1 and cost less: $8 per million input tokens and $40 per million output tokens. An optional fast mode speeds generation by up to 2.5 times, at $16 and $80. The examples are spectacular: a production service migrated from Python to Rust, working electronic devices designed from scratch, and a prototype operating system built by more than 1,000 coordinated agents over eight days.
The number most headlines will skip
Anthropic's own benchmark table does not show Fable 5.2 winning everything. On AutomationBench, a business-workflow test, it scores 51.4%, behind Gemini 4 Carbon's 53.9%. That is no reason to dismiss the model. It is a reminder that the top model overall is not the top model for every job, and that your own tasks are the only benchmark that matters.
Why the maker says "start lower"
The announcement says Fable and Mythos are best for "the most ambitious, open-ended technical work". For most coding and knowledge tasks, Anthropic recommends starting with Opus 5.5 and switching upward or downward, to Sonnet 5.5 or Haiku 5.5, depending on the work.
It also describes two set-ups worth copying. As an Advisor, Fable is consulted by smaller models only when a problem gets hard. As an Orchestrator, it plans the work and hands simpler pieces to smaller models. In both, the expensive model works only where it earns its price.
What that means in money
Here is an illustrative month for a customer-service assistant that answers 10,000 questions, each using about 2,000 input tokens and 500 output tokens. That is 20 million input and 5 million output tokens.
On Fable 5.2 at list price, that is about $160 for input plus $200 for output: around $360 a month. On a smaller model priced at a fraction of that, the same volume could cost a few tens of dollars. If the smaller model answers 95% of questions correctly and only the hard 5% are passed up to Fable, you pay top-model prices on 500 questions instead of 10,000.
The exact numbers depend on the models and prices you choose. The method doesn't: measure the quality you need per task, then pay for no more than that.
What to do with this news
- Keep a test set. Fifty real questions or tasks from your business, with the right answer, re-run whenever a new model comes out.
- Route by difficulty. Routine questions to a small model, hard cases to a large one, with the switch handled in your system, not by hand.
- Watch the safeguards. Fable's restrictions on cybersecurity and biology topics are deliberate. If your work touches those areas, check what your use case needs before you build.
This is how we approach AI projects at SmartHolol: AI agents built on the work your team repeats, with the model chosen per task. For phone lines, Holol-Call puts AI voice agents on inbound and outbound calls, where cost per call handled matters more than any leaderboard.
Bigger is not the strategy
Every few months a more capable model arrives. Companies that can test it on their own tasks in an afternoon, and use it only where it pays, will keep the gains without paying top price for every reply.