Claude Fable 5 launches with 1M context, $10/$50 pricing

Anthropic launched Claude Fable 5 and limited-release Claude Mythos 5 with 1M-token context and new refusal handling.

Anthropic has introduced Claude Fable 5 and Claude Mythos 5, with both models set to become available on June 9, 2026. Fable 5 is the broadly released model; Mythos 5 is limited to approved customers through Project Glasswing.

What changed

Claude Fable 5 is Anthropic’s most capable widely released model, aimed at long-horizon reasoning and agentic workloads. Claude Mythos 5 shares the same capabilities, specs, and pricing, but it is not generally available.

The biggest integration change is that Fable 5 can refuse requests using safety classifiers. In the Messages API, a refusal returns stop_reason: "refusal" with HTTP 200, so teams must treat it as a valid response path rather than an error.

• Fable 5: generally available on the Claude API, Claude Platform on AWS, Amazon Bedrock, Google Cloud, and Microsoft Foundry
• Mythos 5: limited availability through Project Glasswing
• Both models: 1M-token default context window and up to 128k output tokens
• Pricing: $10 per million input tokens and $50 per million output tokens

Anthropic also changed how thinking works. Adaptive thinking is always on, thinking: {"type":"disabled"} is not supported, and developers use the effort parameter to control depth.

Raw chain-of-thought is never returned. Instead, the thinking.display setting can return summarized thinking blocks or omit them entirely, and multi-turn chats must pass those thinking blocks back unchanged when staying on the same model.

Why it matters

For developers, the new refusal flow changes error handling, retry logic, and billing checks. Anthropic says refused requests are not billed before output generation, and fallback retries can recover prompt-cache costs so teams do not pay twice when switching models.

The supported feature set is also broad at launch: effort, task budgets, memory tool, code execution, programmatic tool calling, context-editing result clearing, compaction, and vision. That makes Fable 5 a practical upgrade path for agent builders that need longer context and more controlled reasoning.

Access and compliance are also part of the rollout. Both models carry 30-day data retention and are not available under zero data retention, which matters for teams with stricter storage rules.

The key question for adopters is not whether the models are faster or larger, but how much application logic needs to change to handle refusals, fallback, and hidden reasoning correctly.