Claude Fable 5.2: Use It to Decide, Spend Fewer Tokens

Claude Fable 5.2 is Anthropic's new model for the hardest technical work. The simplest way not to burn tokens is to use it where it really counts: working out what to do, making the calls, writing the plan. Then you hand the building to a model that costs less.

An honest note first: we haven't tried Fable 5.2 in the studio yet. This pill is based on Anthropic's official pages and on a video by Alex Finn, who used it in early access. It's a way of working, not a review.

What It Is

Claude Fable 5.2 came out on 9 November 2026. Anthropic presents it as its most capable model for ambitious, open-ended technical work: coding, reasoning, reading images, and long sessions where an agent works on its own for hours.

It has a twin, Claude Mythos 5.2. According to Anthropic it's the same underlying model, with different cybersecurity and biology safeguards, and only verified organizations can use it. Fable is open to everyone: on the Pro, Max, Ultra, Team and Enterprise plans, and through the API as claude-fable-5-2, including on Amazon Web Services, Google Cloud and Microsoft Foundry.

Why It Matters

Fable 5.2 costs less than the model before it: 8 dollars per million input tokens and 40 per million output tokens, against 10 and 50 for Fable 5.1. Anthropic also says it writes its answers up to 30% faster. It's still the most expensive model in the family. From the API price list, per million tokens:

• Fable 5.2: 8 dollars in, 40 out.
• Opus 5.5: 4 and 20, half the price.
• Sonnet 5.5: 2 and 10, a quarter.
• Haiku 5.5: from 0.10 and 0.50.

There's also a fast mode, up to 2.5 times quicker, that costs double: 16 and 80 dollars.

If you use it for everything, even to rename a file or fix a typo, you pay twice as much for work a cheaper model does well. Anthropic says so itself: for most everyday work it recommends starting with Opus 5.5, stepping up to Fable when the task is really hard, and stepping down to Sonnet or Haiku when it's simple. Fable can also act as an Advisor, the consultant other models turn to on hard problems, or as an Orchestrator, the coordinator that splits the work and passes the simple pieces to smaller models.

How to Use Claude Fable 5.2 Without Burning Tokens

In his video, Alex Finn reaches the same point from practice. In his view Fable 5.2 is at its best in brainstorming and planning: it understands the intent behind a request, not just the words, and it asks a lot of multiple-choice questions, which makes the work less tiring. His workflow, boiled down:

1. Start from an idea, even a vague one. Open Fable 5.2 and explain what you want to achieve, not how to do it. Let it ask questions and answer the choices it offers.
2. Have it write the plan. Once the idea is clear, ask for the phases, the risks, and what "done" means for each part.
3. Turn the plan into tasks. Have Fable write the tasks one by one, each with the criteria that say it's complete, in a tool you'll use later. Finn uses Notion; Linear or a plain file in the project work too.
4. Switch models to build. Move to a cheaper model and have it work through the tasks in order. Finn uses Opus 5.5 at medium effort, meaning less reasoning than the maximum, and says Sonnet would also do.
5. Call Fable back only for decisions. If a task gets stuck or needs an important choice, go back to Fable for that one question, then carry on with the cheaper model.

The prompt for steps 1 to 3, to adapt to your project:

The last line matters most. If the plan leaves the hard choices open, the cheaper model will make them its own way, and the savings go into fixes.

For us the rule isn't new. When we work with Claude Code we already split tasks into critical and simple ones, and the simple ones go to cheaper models: we explain it in the pill on AI coding plans without hitting the limits. With Fable 5.2, only the top step changes, the one where decisions are made.

Small Test

Take a small project you already have in mind, for example a script that cleans up a spreadsheet or a small internal tool. Do only steps 1 to 3 with Fable 5.2 and stop there. Then hand the tasks to Opus 5.5 or Sonnet 5.5.

The test worked if the cheaper model completes the tasks without coming back to ask what you meant. If it keeps asking for clarifications, the problem is the plan: go back to Fable and have it rewrite the plan more precisely.

Watch Out For

• The extraordinary results are Anthropic's claims. The announcement mentions otherwise infeasible software migrations, three electronic devices designed from scratch, two open problems in black hole physics solved, and the proof of Fermat's Last Theorem formalized in 30 hours. These are Anthropic's tests, not independent checks.
• The opinions are Finn's. Smarter, more concise, as fast as Opus: that's his early-access experience, not a measurement.
• Effort shows up on the bill. On the API, Fable 5.2 defaults to high effort, while in Claude Code the default is medium. More effort means more reasoning and more tokens: raise it only when you need to.
• Data is kept for 30 days. For safety monitoring, Anthropic requires 30-day data retention by default. If you work with confidential material, check the terms of your contract.
• Prices and plans change. Check your account and the official Fable page before you decide.

If you want to spend fewer tokens in everyday work too, the next step is the pill on making Claude Code use less. If you need to choose which model to adopt in your company, start with how to choose an AI model with a test on your own process. Want to set up this workflow in your team? Get in touch.

Sources: Anthropic's announcement, the model page and the documentation, all dated 9 November 2026; the video Fable 5.2 just dropped by Alex Finn, checked against the official sources.

In short: Claude Fable 5.2 earns its price when it decides: it writes the plan and the tasks, and a cheaper model does the building.