Claude Fable 5.2 and Mythos 5.2: Access, Pricing and API Migration

Anthropic launched Claude Fable 5.2 and Claude Mythos 5.2 on November 9, 2026. Their most important difference is not a leaderboard result: Anthropic says they share an underlying model, but Fable is generally available while Mythos requires organizational verification for eligible areas such as defensive cybersecurity and life sciences.

For engineering teams the actionable questions are straightforward. Which model can an ordinary API account request? Which old parameters now produce errors? How much does a long-running agent actually cost? And does the model meet data-retention requirements?

The company announcement and official platform release notes establish the November 9 release and model distinctions.

1. Same base model, different access boundary

Anthropic describes Fable 5.2 as generally available across supported Claude applications and APIs. Its Claude API identifier is claude-fable-5-2. Mythos 5.2, identified as claude-mythos-5-2, is available only to approved organizations under applicable verification programs.

Fable includes safeguards for certain cybersecurity and biological research requests. Mythos is offered for verified teams with qualifying applications in those areas. This does not imply that anyone using Claude Enterprise can automatically call Mythos in an arbitrary account. Verify entitlement on the actual platform before updating an application.

Both model documentation pages list:

• 1 million tokens of available context and up to 128,000 output tokens.
• Adaptive thinking always enabled, controlled via supported effort settings.
• A standard price of $8 per million input tokens and $40 per million output tokens in US dollars, before applicable premium modes and other adjustments.
• Text and images as inputs and text output.

See the official Fable model reference and Mythos model reference.

A million-token context window is a published capacity limit, not proof that every detail in a million-token document will be recovered accurately.

2. Lower prices than Fable 5.1 do not mean the lowest-cost model

The preceding Fable and Mythos 5.1 standard rate was $10 input / $50 output per million tokens. The new $8 / $40 rate is a 20% reduction per token against those specific models.

However, Anthropic lists Opus 5.5 at $4 / $20 per million tokens. For identical standard token volumes, Fable 5.2 costs twice as much as Opus 5.5, although a more capable model might reduce retries on some demanding tasks.

For example, illustrative arithmetic, not a measured API workload: 2 million standard input tokens and 400,000 standard output tokens cost $32 using Fable 5.2 (2 × $8 + 0.4 × $40) or $16 with Opus 5.5 (2 × $4 + 0.4 × $20).

Actual billing can differ because of prompt caching, tools, batch jobs, data residency and premium fast mode. Fable’s documented cache-read rate is $0.20 per million cached tokens, with different rates for cache writes. Therefore, cost per accepted task is more useful than a single comparison of input-token prices.

3. Migration-breaking API constraints

Changing only the model ID may not be sufficient. Anthropic’s what’s-new documentation and migration guide describe several concrete issues.

Forced tool choice is not supported

Fable 5.2 rejects forced tool selection modes using tool_choice types any or tool with an API error. Developers relying on model-side forced execution should verify a supported automatic selection configuration and move mandatory authorization or execution rules into the host application.

An agent that must run a security check cannot treat a model’s willingness to choose a tool as an enforcement guarantee.

Sampling fields and thinking behavior differ

The official documentation instructs developers to omit unsupported temperature, top_p and top_k fields. Adaptive thinking is always on; use supported effort controls rather than assuming an older thinking configuration can be replayed unchanged.

Thinking blocks have version- and account-bound restrictions. Agents that persist raw assistant responses, edit earlier turns, or switch models mid-conversation need controlled migration tests, using non-sensitive fixtures.

Retention is a deployment constraint

Anthropic says Fable 5.2 and Mythos 5.2 require 30-day data retention and are not available under zero-data-retention arrangements unless expressly authorized. That may be a harder blocking condition for a production workload than price or benchmark rank.

Review organizational contracts, platform policies and actual deployment requirements before sending private code, customer records or sensitive research material.

4. Performance claims are not a production acceptance test

Anthropic reports improved coding, reasoning and scientific work, including demanding prerelease projects, and describes safety testing with internal and external evaluators. Those are useful disclosure points, but not an independent benchmark run by this publication.

Fable’s safeguards can block some categories of requests or route them to other models. Such behavior matters to regulated or technical workflows, but one refusal-rate claim cannot predict the outcome of every request.

The question for an engineering manager is whether the model produces more accepted results per dollar and review hour on the organization’s own task set. Measure successful final artifacts, actual API costs, elapsed time, repeated tool calls and whether the evidence still applies to the final edited files.

Our MCP test-evidence investigation explains why a passing test command and verified final code are distinct claims. The Cline/Roo Code/Continue comparison provides additional client-side permission context.

5. Practical migration checklist

1. Confirm the API provider and actual model access; Mythos requires additional approval.
2. Pin and log the model ID alongside the SDK and platform version.
3. Validate tool-choice, sampling, thinking and saved-conversation behavior in a staging environment.
4. Confirm 30-day retention is compatible with the application’s obligations.
5. Re-run identical tasks under identical tool permissions. Record cost per accepted outcome, not just token prices.
6. Retain a rollback path to a known-good model and an enforceable tool-execution policy.

Bottom line: Fable 5.2 is a generally available release and Mythos 5.2 is a restricted counterpart sharing its underlying model. Access, migration behavior, data handling and total task cost matter at least as much as Anthropic’s published capability claims.

Primary sources

• Anthropic: Fable 5.2 and Mythos 5.2 launch, November 9, 2026
• Fable 5.2 model specification and pricing
• Mythos 5.2 availability and requirements
• Fable 5.2 changes and compatibility guidance
• Fable 5.2 migration documentation

Editorial note: Published November 11, 2026, based on the November 9 announcement and official API documents. Agentic Pulse did not receive Mythos access, run a model benchmark or accept compensation or editorial direction from Anthropic.