Anthropic launched Claude Fable 5.2 and Claude Mythos 5.2 on November 9 for technical work that spans several stages of execution and verification. On FrontierCode 2.0, Fable 5.2 nearly matched Opus 5.5 at about half the cost per task, even though its list price per token is twice as high. The comparison depends on the effort settings and illustrates why the cost of a completed task is the more useful unit.

What separates Fable and Mythos

Both use the same underlying model. Fable is broadly available with safeguards for high-risk cybersecurity and biology work. Mythos has fewer restrictions in those areas and is available to organizations approved through Anthropic's verification programs. Enterprise customers can also use Mythos indirectly through Claude Security's repository scans.

Both models have a one-million-token context window, 128,000 maximum output tokens, and always-on adaptive thinking. Default effort is high on the API and medium in Claude Code and the Claude apps. Anthropic reports output generation up to 30% faster than the 5.1 models. Optional fast mode generates tokens up to 2.5 times as fast. Total runtime will also depend on how long a workflow spends running tests or waiting for tools. Anthropic's launch announcement.

Availability, pricing, and usage resets

Fable 5.2 is available in Claude Code and Claude apps for eligible subscribers and in the Claude API as claude-fable-5-2. Anthropic's documentation lists Amazon Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry. Mythos uses claude-mythos-5-2 on the Claude API. Organizations seeking direct access can apply through the cybersecurity or life sciences programs or contact their account team.

Model or mode

Fable 5.2 and Mythos 5.2

Input per million tokens
$8
Output per million tokens
$40
Model or mode

Fable 5.1 and Mythos 5.1

Input per million tokens
$10
Output per million tokens
$50
Model or mode

Opus 5.5

Input per million tokens
$4
Output per million tokens
$20
Model or mode

Fable 5.2 and Mythos 5.2 fast mode

Input per million tokens
$16
Output per million tokens
$80

Cache reads cost $0.20 per million tokens for the 5.2 models, and batch processing reduces base input and output prices by 50%. Claude Platform documentation.

  • Pro and standard Team seats can spend up to half of their weekly usage allowance on Fable models at no extra charge through November 16 at 11:59 p.m. Pacific. Eligible accounts receive one reset with the same expiry. After that window or after the allowance is exhausted, Fable usage depends on enabled usage credits.
  • Max, Ultra, premium Team seats, and premium seats on legacy seat-based Enterprise plans receive one to three resets, depending on subscription tenure. These expire March 8, 2027, at 11:59 p.m. Pacific.
  • Settings → Usage is where users can restore consumed Fable usage. The general paid billing cutoff is November 9 at 10 a.m. Pacific, with specific rules for trials and organizational seats. Claude Help Center.

Coding, automation, and vision benchmarks

Anthropic reports each model's highest available effort unless otherwise indicated. On FrontierCode, both Fable versions used xhigh, while Gemini used medium. OSWorld uses an offline subset with partial credit, and Chartography tests chart reading without tools.

Benchmark

FrontierCode 2.0

Fable 5.2
49.1%
Fable 5.1
32.4%
Opus 5.5
45.5%
GPT-6 Astra
42.8%
Gemini 4 Carbon
41.1%
Benchmark

FrontierSWE v2

Fable 5.2
69.9%
Fable 5.1
56.3%
Opus 5.5
62.3%
GPT-6 Astra
65.5%
Gemini 4 Carbon
61.2%
Benchmark

AutomationBench

Fable 5.2
51.4%
Fable 5.1
31.4%
Opus 5.5
42.5%
GPT-6 Astra
41.4%
Gemini 4 Carbon
53.9%
Benchmark

Terminal-Bench-Science 0.1

Fable 5.2
74.6%
Fable 5.1
52.6%
Opus 5.5
58.7%
GPT-6 Astra
68.1%
Gemini 4 Carbon
64.8%
Benchmark

OSWorld 2.1

Fable 5.2
89.3%
Fable 5.1
83.5%
Opus 5.5
87.2%
GPT-6 Astra
79.5%
Gemini 4 Carbon
81.5%
Benchmark

Chartography

Fable 5.2
73.4%
Fable 5.1
44.8%
Opus 5.5
64.4%
GPT-6 Astra
71.0%
Gemini 4 Carbon
72.5%

On FrontierCode, which evaluates code changes in open-source projects, Fable 5.2 at high scored 45.3% against Opus 5.5's 45.5% at max, for roughly half the cost per task. Fable reached 49.1% at xhigh. Gemini 4 Carbon led on AutomationBench, a business workflow evaluation. The results apply to the reported test settings. Launch benchmarks and methodology.

Mythos 5.2 scored 89.6% on CVE Rediscovery, compared with 78.3% for Mythos 5.1 and 81.3% for Opus 5.5 with safeguards disabled. Anthropic's internal test evaluated rediscovery of 1,829 published vulnerabilities on code snapshots from before the fixes. The Mythos score reflects its configuration and should be read alongside the safeguards and access differences.

Using Fable to coordinate Opus on a migration

As an orchestrator, Fable can decompose a project and delegate work. As an Advisor, Fable can answer questions from an executor such as Opus, Sonnet, or Haiku when a difficult decision comes up. Anthropic recommends starting with Opus 5.5 for most coding tasks and moving to Fable when evaluation results justify it.

On a legacy migration, Fable could map dependencies and define the contracts to preserve. Opus agents could implement isolated changes and return diffs and test results. Fable could then integrate the outputs and investigate mismatches across modules. The extra coordination pays off when it reduces rework or gets the team to an accepted result sooner.

Anthropic reports that agents using a prerelease version helped migrate a critical service from Python to Rust. The new service started about 26 times faster. That figure measures the resulting software. Evaluating the migration process itself would take engineering time, regression, and execution cost data.

In mathematics, roughly 1,000 Fable 5.2 agents took 30 hours to formalize the known proof of Fermat's Last Theorem in Lean. In the same setup, Opus 5.5 reached 26% in 18 hours and made little further progress over the next 12 hours. There was one run per model, on a task with a precise verification criterion. Anthropic's reported experiments.

What changes for API integrations

  • Editing earlier messages, instructions, or tools can invalidate preserved thinking blocks.
  • When switching from Fable 5.2 to Opus 5.5, the API drops thinking blocks that Opus cannot read. Diffs, test logs, and recorded decisions can carry context between agents.
  • Forced tool use with tool_choice: any or tool returns an error. Computer-use integrations should check the new toolset and platform exceptions.
  • The models require 30-day data retention unless Anthropic expressly authorizes zero data retention. Claude Platform migration details.

A useful comparison runs the same difficult task with Opus alone, Fable alone, and Fable coordinating Opus, using the same acceptance tests. Cost per approved delivery includes review time. For migrations, the outcome that matters is whether the new service preserves the behavior users already rely on.