Claude Fable 5 is live for developers as of June 9, and the integration surface has a few behaviors you will not know from Opus or Sonnet. This guide collects the launch-day facts that matter when you wire it in: IDs, limits, the new refusal semantics, and where the model actually runs.
The basics: IDs, context, output
- API model ID: claude-fable-5. The restricted sibling is claude-mythos-5, Glasswing-only.
- Context window: 1M tokens by default, with up to 128k output tokens per request.
- Pricing: $10 per million input tokens, $50 per million output. Exactly double Opus 4.8.
All of this is documented in Anthropic's launch notes for the platform, which also answer the context-window question that was still open on launch morning.
Adaptive thinking is always on
Fable 5 has exactly one thinking mode: adaptive. You cannot disable it; thinking type disabled is not supported on this model. Depth is steered through the effort parameter instead. One more change worth planning around: the raw chain of thought is never returned. Thinking display defaults to omitted, and if you want something readable you set display to summarized and get a condensed version. If your tooling parses thinking blocks from other Claude models, it needs this branch.
Refusals are a 200, not an error
The most consequential integration change sits in the safety layer. When Fable 5's classifiers decline a request, the Messages API returns a successful HTTP 200 with stop_reason: "refusal", and the response tells you which classifier fired. Error-based retry logic will sail right past it.
Anthropic ships two handling paths:
- Server-side: pass the fallbacks parameter (beta on the Claude API and the Claude Platform on AWS) and the API retries on another model for you.
- Client-side: SDK middleware for TypeScript, Python, Go, Java, and C# retries from your side on any platform.
Billing is friendlier than expected: a request refused before any output costs nothing, and when you retry on another model, fallback credit refunds the prompt-cache cost of the switch. The pricing background and the June 22 subscription window are covered in our access guide.
Launch-day feature set
Supported at launch: the effort parameter, task budgets (beta, header task-budgets-2026-03-13), the memory tool, tool result clearing through context editing (beta header), compaction, and vision. The memory and compaction pieces are what make the long-horizon claims practical; we look at that side in our agents deep dive.
Where it runs
- Claude API plus the Claude Platform on AWS: generally available now.
- Amazon Bedrock, Vertex AI, Microsoft Foundry: generally available now.
- GitHub Copilot: rolling out to Pro+, Max, Business, and Enterprise, across VS Code (chat, ask, edit, and agent modes), Visual Studio, JetBrains, Xcode, Eclipse, the Copilot CLI, the cloud agent, github.com, and GitHub Mobile, per the GitHub changelog. Rollout is gradual.
One admin detail with teeth: on Copilot Business and Enterprise, Fable 5 is disabled by default. An administrator has to enable it and acknowledge that Anthropic retains prompts and outputs for up to 30 days to operate the safety classifiers. If you are on a team plan and cannot see the model, that toggle is why.
Migrating from Opus 4.8
Anthropic publishes a step-by-step migration guide, and the short version is: change the model ID, remove any thinking-disable logic, adopt effort, and add refusal handling. Budget-wise, runs finish in fewer turns, which claws back part of the 2x token price; the math is in our benchmark and cost comparison. For what the model is and why it exists in two versions, start with the Fable 5 explainer.