Claude Fable 5 is live again for everyone. As of July 1, 2026, Anthropic's most capable widely released model is generally available in Claude.ai, Claude Code, and the API (model ID claude-fable-5), after an 18-day outage tied to US export controls. Fable 5 is not a creative or video tool. Per Anthropic's own docs, it is "built for the most demanding reasoning and long-horizon agentic work," a step above Opus 4.8, priced at $10 per million input tokens and $50 per million output. It also ships one change every builder has to handle: Fable 5 can now decline requests, and how you deal with that refusal decides whether your agent keeps running.
What Claude Fable 5 Actually Is
Fable 5 sits at the top of Anthropic's lineup as its most capable generally available model, aimed at hard reasoning and long-running, asynchronous agentic tasks, the kind of work you hand off and let run. It carries a 1M-token context window and returns up to 128,000 output tokens per request. Its sibling, Claude Mythos 5, shares the exact same capabilities and specs but ships without safety classifiers and is only offered to approved customers through Project Glasswing. For everyone else, Fable 5 is the model you actually get.
Fable 5 vs Opus 4.8 vs Mythos 5
Fable 5 is positioned above Opus 4.8, and priced accordingly. Here is how the three line up. All prices are per million tokens (MTok).
| Spec | Fable 5 | Opus 4.8 | Mythos 5 |
|---|---|---|---|
| Best for | Most demanding reasoning + long-horizon agents | Top-tier reasoning flagship | Same as Fable 5, unfiltered |
| Input / output | $10 / $50 | $5 / $25 | $10 / $50 |
| Context / max output | 1M / 128K | 1M / 128K | 1M / 128K |
| Safety classifiers | Yes (can refuse) | Standard | None |
| Availability | Generally available | Generally available | Project Glasswing only |
The takeaway: Fable 5 costs twice Opus 4.8 per token, so you reach for it when a task genuinely needs the extra capability and can run long, not for routine work. For cost-sensitive default work, the cheaper tiers still win, as we covered in our Claude Sonnet 5 breakdown.
How to Access Fable 5 Right Now
Getting to Fable 5 depends on how you build. General users simply select it in Claude.ai or use it in Claude Code, with no early-access scheme required. Developers call claude-fable-5 on the Claude API, and it is also generally available on the Claude Platform on AWS, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry.
Watch the pricing calendar closely. Through July 7, Anthropic includes Fable 5 for up to 50% of weekly usage limits on Pro, Max, Team, and select Enterprise plans, so heavy exploration is effectively bundled this week. After July 7, access moves to metered usage credits at the standard $10 and $50 per million tokens, with a 90% input discount when you use prompt caching. If you plan to lean on Fable 5, front-load your testing before the window closes and confirm current terms on the Claude pricing page.
The Big Change: Refusals, Fallback, and Billing
This is the part that breaks naive integrations. Fable 5 includes safety classifiers that can decline certain requests. When it refuses, the Messages API returns stop_reason: "refusal" as a successful HTTP 200 response, not an error, and it reports which classifier declined. Code that only checks for HTTP errors will treat a refusal as a valid, empty answer and quietly break.
Anthropic pairs that with a fallback system. A request Fable 5 refuses can usually be served by another Claude model, and there are three ways to retry: server-side via the fallbacks parameter (in beta), client-side through the official SDK middleware for TypeScript, Python, Go, Java, and C#, or a manual retry you build yourself. On billing, you are not charged for a request that is refused before any output is generated, and when you retry on another model, fallback credit refunds the prompt-cache cost of switching so you do not pay it twice.
How to Add Fable 5 to Your Stack Safely
If you are wiring Fable 5 into an agent or product, do it in this order so a refusal never becomes a silent failure.
- Pin the model deliberately. Set
claude-fable-5only on the tasks that need top-tier reasoning or long-horizon autonomy. Keep cheaper models on routine calls. - Handle the refusal stop reason. Check for
stop_reason: "refusal"on every response and branch on it, rather than assuming a 200 means a usable answer. - Wire a fallback before you ship. Use the server-side
fallbacksparameter or the SDK middleware so a refused request retries on another Claude model automatically. - Budget for credits and caching. After July 7 you pay metered credits, so estimate token spend, turn on prompt caching for the 90% input discount, and set a ceiling.
- Tune effort, not thinking mode. Adaptive thinking is always on and cannot be disabled; use the
effortparameter to control depth and cost instead. - Test long-running jobs first. Fable 5 is built for asynchronous, long-horizon work, so validate a real multi-step run end to end before trusting it in production.
Capabilities and Behaviors to Know
A few Fable-specific behaviors change how you build:
- Adaptive thinking is the only mode. It is always on;
thinking: {"type": "disabled"}is not supported. Control depth with the effort parameter. - Raw chain of thought is never returned. Thinking blocks are either a readable summary or omitted (the default); pass them back unchanged in multi-turn conversations on the same model.
- Tooling is broad at launch: effort, task budgets (beta), the memory tool, code execution, programmatic tool calling, context editing, compaction, and vision are all supported.
- Data retention is 30 days. Fable 5 and Mythos 5 are Covered Models and are not available under zero data retention.
- Migration path exists. Anthropic publishes a step-by-step migration guide for moving from Opus 4.8 to Fable 5.
Frequently Asked Questions
Is Claude Fable 5 available now?
Yes. As of July 1, 2026, Fable 5 is generally available in Claude.ai, Claude Code, and the API (model ID claude-fable-5), plus AWS, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry.
Is Fable 5 a video or creative model?
No. Despite the name, Anthropic describes Fable 5 as its most capable widely released model, built for demanding reasoning and long-horizon agentic work, not image, video, or audio generation.
How much does Fable 5 cost?
$10 per million input tokens and $50 per million output tokens, with a 90% input discount when using prompt caching. Through July 7, it is included for up to 50% of weekly usage limits on Pro, Max, Team, and select Enterprise plans; after that it is metered usage credits.
What happens when Fable 5 refuses a request?
The API returns stop_reason: "refusal" as a successful HTTP 200 response and reports which classifier declined. You are not billed for a refusal before output, and you can retry on another Claude model via server-side, client-side, or manual fallback.
How is Fable 5 different from Mythos 5?
They share the same capabilities and specs, but Mythos 5 ships without safety classifiers and is only available to approved customers through Project Glasswing. Fable 5 is the generally available model with classifiers that can refuse.
Can I turn off thinking on Fable 5?
No. Adaptive thinking is always on and cannot be disabled. Use the effort parameter to control how much thinking, and therefore cost, each request uses.
