About Fable 5.2, the new model from Anthropic

Anthropic has released Claude Fable 5.2, its most capable widely released model. It's built for the hardest reasoning problems and for long-horizon agentic work: tasks where a model plans, uses tools and keeps going for many steps before it hands back a result.

We spend a lot of time choosing models for the systems we build, so here is our read on what Fable 5.2 is, what's new, and where it fits next to the rest of the Claude family.

What Fable 5.2 is

Fable sits above the Opus, Sonnet and Haiku lines. Where Opus is the strong default for most serious work, Fable is the model you reach for when the problem is genuinely hard and getting it right matters more than speed or cost.

The essentials:

Model ID: claude-fable-5-2
Context window: 1 million tokens, which is also the default
Maximum output: 128K tokens in a single response
Price: $8 per million input tokens and $40 per million output tokens, with cached input read at $0.20 per million

It replaces Claude Fable 5.1, which is still served, at a lower price: Fable 5.1 costs $10 and $50 per million tokens.

Reasoning that is always on

Fable 5.2 always thinks before it answers. There is no switch to turn reasoning off, and fixed thinking budgets are gone. Instead you choose an effort level, from low and medium through high, xhigh and max, which sets how deeply the model works on each request.

Anthropic's guidance is to start at high. The gains above it are uneven, so xhigh is worth it only where you can measure an improvement, and max is best kept for finding the model's ceiling. Routine work often does well at low or medium.

The raw chain of thought is never returned. If you want to show users what the model is doing while it works, you can ask for a readable summary of its reasoning instead.

Built for long tasks

Because it is designed for long, demanding work, a single request on a hard task can run for many minutes. That has practical consequences:

Stream responses instead of waiting for one large reply, so requests don't hit timeouts.
Show progress in your product, so a long turn doesn't look like a frozen screen.
Rewrite old prompts. Instructions written for earlier models are often too prescriptive and can make the output worse. Give Fable the full task and the result you want, and let it plan.

For teams that need it, fast mode runs the same model at up to 2.5 times the output speed, at $16 per million input tokens and $80 per million output tokens.

How it compares

Model	Best for	Input / output per 1M tokens
Claude Fable 5.2	The hardest reasoning and long agentic tasks	$8 / $40
Claude Opus 5.5	The default for serious, complex work	$4 / $20
Claude Sonnet 5.5	Everyday coding, agents and enterprise work	$2 / $10
Claude Haiku 5.5	Fast, high-volume tasks	$0.10 / $0.50

Every one of these models now offers a 1M-token context window, so the choice comes down to how hard the task is and how much each completed task is allowed to cost.

Things to know before you switch

Data retention: Fable 5.2 requires 30-day data retention. It isn't available under zero data retention unless Anthropic has expressly authorized it.
Refusals: safety classifiers can decline a request. Your code should check for a refusal stop reason, and Anthropic offers server-side fallbacks that rerun a declined request on another model inside the same call.
A sibling model: Claude Mythos 5.2 has the same capabilities, price and API, and is available to verified organizations only.

Our take

Fable 5.2 isn't the model for every request, and it doesn't try to be. For most products, Opus 5.5 or Sonnet 5.5 at the right effort level will do the job for less. But for the work where a weaker answer is expensive (a complex migration, an agent that runs for hours, analysis across a whole codebase or document set) a cheaper and more capable Fable is a welcome change.

Our advice is the same as always: measure on your own tasks, judge cost per completed task rather than per request, and move up a model only when the results earn it.