Fable 5.2 for business: what changed, what it costs and when to use it

Not for everything. Anthropic itself recommends starting most coding and knowledge work on Opus 5.5. Use Fable 5.2 for big, open-ended jobs such as root-cause investigations, large refactors and architecture decisions. If you are on Pro or a standard Team seat, test it before 16 November. Until then it is included in your weekly limits.

Anthropic announced Claude Fable 5.2 and Claude Mythos 5.2 on 9 November 2026. Fable 5.2 generates outputs up to 30% faster than Fable 5.1. That is Anthropic's own figure. An optional fast mode speeds up token generation by up to 2.5 times.

The standard price is $8 per million input tokens and $40 per million output tokens. Fast mode costs $16 per million input tokens and $80 per million output tokens. So fast mode doubles both prices. The Claude Code changelog adds that cache reads cost $0.20 per million tokens.

Solid: standard price Hatched: fast mode

Anthropic says Fable 5.2 is available on all platforms. That includes Amazon Web Services, Google Cloud and Microsoft Azure. The API model ID is claude-fable-5-2. In Claude Code it has a 1M token context window. It needs Claude Code version 2.1.321 or later.

Mythos 5.2 is the same generation. It is only for vetted organisations through Anthropic's verification programmes. So in practice I expect most teams to compare Fable 5.2 with Opus 5.5.

The launch offer depends on your plan. On Pro and standard Team seats, all Fable models are included in the plan's weekly usage limits until 16 November 2026. Fable can use up to 50% of your weekly limits in that time. After 16 November, Fable runs on usage credits. That only happens if you have turned credits on. Unused allowance does not carry over.

Anthropic also gives Fable resets. Pro and standard Team seats get one. It expires on 16 November 2026 at 11:59 pm PT. Max, Ultra, premium Team and premium legacy Enterprise seats get one to three. The number depends on how long you have subscribed. These resets expire on 8 March 2027 at 11:59 pm PT.

Scroll sideways to see every column

There is a cutoff for resets. They only go to accounts with paid billing in place before 9 November 2026 at 10:00 am PT.

On the API you pay per token at the prices in section 01. Cloud sessions in Claude Code count toward your plan's usage limits. There is no separate charge for the cloud machine.

Cost is one side of the sum. The other side is what AI brings in. My note on how analytics hides AI search revenue shows one way to measure that for customers who find you through AI assistants.

Anthropic reports two coding benchmarks. FrontierCode 2.0 tests agentic coding. On it, Fable 5.2 scored 49.1% at xhigh effort. At high effort it scored 45.3%. Opus 5.5's best was 45.5%. On FrontierSWE v2, Fable 5.2 scored 69.9% and Opus 5.5 scored 62.3%.

Hatched: Fable 5.2 Solid: Opus 5.5

My view: these are Anthropic's own tests of its own models. At high effort, Fable 5.2 is level with Opus 5.5 on FrontierCode. The clear lead there only appears at xhigh effort. Both models default to medium effort in Claude Code. So the scores above are not at the setting most people will use.

The speed figures say nothing about quality. Faster output only helps when the work is right. "Up to 30% faster" is also a ceiling. Your own jobs may see less.

My note on how to read an AI benchmark claim sets out the questions I put to any vendor's numbers. The best evidence is still a small trial on your own work. Section 06 shows one.

Anthropic says Fable and Mythos are best for the most ambitious, open-ended technical work. For most coding and knowledge tasks it recommends starting with Opus 5.5. It calls Opus 5.5 its strongest daily driver.

The Claude Code docs say Fable models suit tasks larger than a single sitting. They sustain long sessions. They investigate before acting. They also check their own work more often than smaller models.

Scroll sideways to see every column

Anthropic also gives tips for working with Fable. Describe the outcome you want and leave the steps to the model. Hand it ambiguous problems such as root-cause investigations and architecture decisions. You do not need to remind it to check its work. Give it the work you would normally split into several tasks.

To switch in Claude Code, run /model fable. You need version 2.1.321 or later. Later you may switch to a model with a smaller context window. If the conversation is too long for it, Claude Code now has the previous model summarise it first.

Models can differ in ways a benchmark does not show. My note on how ChatGPT, Claude, Gemini and Perplexity answer the same question shows how far assistants can disagree. That is one more reason to test more than one model on the same task.

The Claude Code changelog lists several recent changes that matter for teams. I would set these up before a trial.

• Block writes outside the project. The setting permissions.blockWritesOutsideWorkingDirectories makes Write, Edit and NotebookEdit refuse paths outside the working directories. It applies in every permission mode.
• Set up auto mode with care. /auto-mode-setup is a guided setup. It reads your project and recent sessions. Then it teaches auto mode which repos and services you trust.
• Keep API keys away from the model. On a cloud environment you can store an API key or token as a network secret. Claude can then call that API without seeing the key. Network secrets are available on Pro, Max, Ultra, Team and Enterprise.
• Check your usage. /usage now shows each warning as a one-line fix that names its source. Press f to apply the fixes.
• Leave idle compaction on. A conversation over 200K tokens left idle for about 54 minutes is compacted before its one-hour prompt cache expires. So the next message does not reprocess it all. You can turn it off in /config.
• Keep ultrathink for stuck problems. ultrathink runs one turn at xhigh effort. My view is to use it only when a problem is stuck.

Each of these limits what an agent can reach or spend. My note on AI agents and the Model Context Protocol explains why limits on access matter for any agent.

Pick three real tasks. Run each one on both models with the same prompt. For each run, note the time to finish, the rework needed and the share of your weekly limit used. If you are on Pro or a standard Team seat, do the Fable runs first. They are only included in your weekly limits until 16 November.

Free template: three tasks

1. A root-cause investigation on a bug you have already fixed. You know the answer, so you can judge the result.
2. A routine edit your team makes every week.
3. A refactor of one module that you would normally split into several tasks.

Free template: the scoresheet

Record the results in a table like this one. The example below is filled in for an imaginary team so you can see how it reads.

Scroll sideways to see every column

In this made-up example, Fable 5.2 wins the investigation and the refactor. Opus 5.5 does the routine edit just as well for a third of the limit. Use your own tasks and your own numbers. A result on your own work tells you more than any launch table.

Some teams want help to choose models and to set rules for how their people use them. If that is your team, see work with me. It starts with a free 30-minute discovery call.