# Claude Fable 5.2: Anthropic's model for demanding reasoning and multi-hour agents

> What Claude Fable 5.2 is, which benchmarks Anthropic published and how to read them, what it costs ($8/$40) and when to pick it. Data as of November 15, 2026.

*Tags: ai, models, anthropic, reasoning, agents*

---


> [!info] Quick Definition
> **Claude Fable 5.2** is a closed language model from Anthropic, released on November 9, 2026, built for demanding reasoning and long-running agent work. It is the most expensive of the four most recent Claude models (Fable 5.2, Opus 5.5, Sonnet 5.5 and Haiku 5.5), and it shares its base model with Claude Mythos 5.2, which is reserved for vetted organizations.


## What is Claude Fable 5.2 and when did it come out?

It is the latest version of the Fable line, part of a Claude family that also includes [[claude-opus-5-5|Opus 5.5]], Sonnet 5.5 and [[claude-haiku-5-5|Haiku 5.5]]. Anthropic says its output is up to 30% faster than Fable 5.1. It thinks before it answers, using adaptive thinking that cannot be switched off (see [[reasoning-models|reasoning models]]).

| Fact | Value |
|---|---|
| API ID | `claude-fable-5-2` |
| Release | November 9, 2026 |
| Context and output | 1M tokens of context; up to 128K of output |
| Price | $8 input / $40 output, per million tokens |
| Input and output | Text and images in, text out |
| Knowledge cutoff | June 2026 |
| Open or closed | Closed: used through the API and cloud platforms |

Fable and Mythos are the same model with different safeguards. If you are not in a verification program, you use Fable.

## What is it for?

Anthropic describes it for agent sessions that run for hours, multistep deep research, and analysis carried through to a finished document, spreadsheet or deck. Anthropic itself recommends starting with Opus 5.5 for most coding and knowledge tasks, and moving up to Fable for the most ambitious work or when your tests with Opus 5.5 at higher effort still fall short.

For a designer, that means long jobs with many pieces: cross-referencing dozens of [[claude-for-designers|research]] documents, preparing a full report, or leaving an agent working in [[claude-code|Claude Code]] while you do something else.

Three usage details matter if you call the API, according to the documentation: forcing a tool (`tool_choice` of type `any` or `tool`) returns a 400 error; you must omit `temperature`, `top_p` and `top_k`; and it requires 30-day data retention.

## Which benchmarks did Anthropic publish and how do you read them?

Anthropic reports these scores in its announcement, with safeguards on and, unless noted, at each model's highest available effort; the methodology is in the system card. They are lab data, not an independent measurement.

| Benchmark | Score | Reported by | What it measures, per the source |
|---|---|---|---|
| FrontierCode 2.0 | 49.1% (Xhigh effort) | Anthropic | Whether code changes would be merged by open-source maintainers |
| FrontierSWE v2 | 69.9% | Anthropic | Agentic coding; the page does not describe it |
| Terminal-Bench-Science 0.1 | 74.6% | Anthropic | Agentic scientific research (category) |
| OSWorld 2.1 (offline set, partial) | 89.3% | Anthropic | Operating a real computer on long, multi-step tasks |
| Humanity's Last Exam, with tools | 74.2% | Anthropic | Expert-level academic knowledge and reasoning |
| Remote Labor Index | 41.3% (High effort) | Anthropic | Knowledge work; the page does not describe it |

How to read them:

- **Check the full name and the configuration.** FrontierCode 2.0 was run at Xhigh effort and the Remote Labor Index at High. With a different effort, the score changes: Anthropic reports 45.3% on FrontierCode at High effort, versus 49.1% at Xhigh.
- **Compare only within the same lab and the same table.** In that table, Fable 5.1 scored 32.4% on FrontierCode 2.0 and Fable 5.2 scored 49.1%. That comparison holds because Anthropic ran both on the same test.
- **Do not build rankings against models from other labs.** Each lab runs its tests its own way.
- **A score is not your case.** To know how it behaves on your tasks, use your own evaluation; the guide is in [[ai-benchmarks|how to read AI benchmarks]].

## How much does it cost and how much context does it have?

As of November 15, 2026, Anthropic's documentation lists these prices per million tokens:

- Input: $8. Output: $40.
- Cache read: $0.20. 5-minute cache write: $10.
- Batch processing (Batch API): 50% discount on input and output.
- Fast mode (up to 2.5 times faster generation): $16 input / $80 output.

Fable 5.1 is listed at $10 / $50, so 5.2 costs less. The 1M-token context is about 555,000 words, according to the documentation. If you are unsure what a token is, see [[tokens|tokens]] and [[cost-per-token|cost per token]].

Illustrative example of the math: a task with 200,000 input tokens and 20,000 output tokens costs $1.60 + $0.80 = $2.40. Real spend depends on how much the model produces; with [[prompt-caching|prompt caching]] you can lower the cost of whatever you repeat.

## When should you pick it and when not?

Pick it when:

- A hard task already failed with [[claude-opus-5-5|Opus 5.5]] at higher effort.
- The work is long, with many steps, and the cost of a mistake is higher than the cost of the call.
- You need a finished deliverable (report, spreadsheet, deck) built from a lot of material.

Do not pick it when:

- The task is high-volume and low-complexity: classifying, extracting or routing. That is where [[claude-haiku-5-5|Haiku 5.5]] fits, at a fraction of the price.
- Your organization requires zero data retention.
- You need to control temperature or force a tool call.
- You have not yet measured whether Opus 5.5 is enough. That is what Anthropic recommends trying first.

To see all the models side by side, check [[current-ai-models|current AI models]].

## Example

Illustrative example: a designer has 40 anonymized interview transcripts and needs a findings report with quotes and a map of problems per screen. She first runs the task on Opus 5.5 at higher effort and saves the result. She checks ten quotes by hand against the original and notes how many fail. If the report misses themes or invents quotes, she repeats exactly the same task with Fable 5.2, compares both results against the same written criteria, and works out the cost of each run.

What to try tomorrow: pick a real task that currently takes you an afternoon, write three success criteria and run the same task on both models. Buy Fable only if the difference justifies the price.

## Common mistakes

- Starting with the most expensive model without first testing the previous one at higher effort.
- Treating Anthropic's scores as a ranking against models from other labs.
- Budgeting with one month's price and never checking again: prices change.
- Passing `temperature` or forcing tools and getting a 400 error.
- Uploading user data without checking the retention policy.
- Accepting quotes and reports without verifying them against the source.

## Sources

- [Anthropic: introducing Claude Fable 5.2 and Claude Mythos 5.2](https://www.anthropic.com/claude-fable-and-mythos-5-2)
- [Documentation: Claude Fable 5.2](https://platform.claude.com/docs/en/models/fable-5-2/overview)
- [Documentation: Claude models overview](https://platform.claude.com/docs/en/about-claude/models/overview)
- [Documentation: pricing](https://platform.claude.com/docs/en/about-claude/pricing)
- [Anthropic: introducing Claude Haiku 5.5 (descriptions of OSWorld 2.1 and Humanity's Last Exam)](https://www.anthropic.com/news/claude-haiku-5-5)


---

Source: https://www.fernandoux.com/en/wiki/ai/claude-fable-5-2/
