What Is Claude Fable 5? Anthropic's Most Capable Model

Claude Fable 5 is Anthropic's most capable widely released model, built for the most demanding reasoning and long-horizon agentic work and available since June 9, 2026. It uses the API model ID claude-fable-5, ships with a 1M-token context window by default, and is priced at $10 per million input tokens and $50 per million output tokens.

Update — June 23, 2026: Claude Fable 5 and Claude Mythos 5 are currently suspended and unavailable. Anthropic disabled both models worldwide within roughly 72 hours of the June 9 launch after a U.S. export-control directive, while Claude Opus 4.8, Sonnet 4.6, and Haiku 4.5 stayed online. Reports describe this as a temporary suspension, not a permanent ban.

What Claude Fable 5 is

Claude Fable 5 is the top tier of Anthropic's lineup, positioned for the most demanding reasoning and long-horizon agentic work rather than everyday chat. It became available on June 9, 2026, and is exposed through the API model ID claude-fable-5. Where lighter models trade depth for speed and cost, Fable 5 is the model you reach for when a task spans many steps, large codebases, or research that has to stay coherent across a long session.

Anthropic describes Fable 5 as its most capable model, and the design choices reflect that. It uses the same tokenizer as Claude Opus 4.8, so token counts are roughly unchanged if you are migrating from Opus 4.7 or 4.8. Anthropic has not published specific Fable 5 benchmark scores at launch, so the honest framing is qualitative: it is built to handle the hardest reasoning and agentic workloads in the family. You can read the full feature set on the official Claude Fable 5 introduction.

Context window, output, and pricing

Claude Fable 5 ships with a 1M-token context window by default and can return up to 128K output tokens in a single request. Pricing is $10 per million input tokens and $50 per million output tokens, which makes it the premium option in the lineup. The table below compares it directly to Claude Opus 4.8 so you can see exactly where the premium sits.

The headline difference is price, not size: both models share the same 1M context and 128K output ceiling, but Fable 5 costs roughly double per token. That gap matters most on output-heavy workloads, since output is the expensive side at $50 per million. For a model-by-model breakdown of how the rest of the family lines up, the Anthropic models overview is the canonical reference.

What changes for developers

Adaptive thinking is always on in Claude Fable 5, and it is the only thinking mode available. You cannot pass thinking:{type:"disabled"} the way you might on other models; instead you control how hard the model thinks with the effort parameter, which accepts low, medium, high, xhigh, and max. The effort documentation covers how each level trades depth against cost and latency.

The raw chain of thought is never returned. Thinking blocks come back either as a summary (displayed as "summarized") or empty (displayed as "omitted", which is the default), so do not build logic that depends on parsing the model's internal reasoning. Fable 5 also drops several familiar knobs: there is no assistant prefill, and temperature, top_p, and top_k have been removed.

Fable 5 includes safety classifiers that can decline a request. When that happens, the Messages API returns stop_reason:"refusal" as a successful HTTP 200, not an error, and you are not billed for a request refused before any output. Plan for it: you can retry on another Claude model using the server-side fallbacks parameter, client-side SDK middleware, or a manual fallback. One operational constraint to flag early is data handling: Fable 5 requires 30-day data retention and is designated a Covered Model, so it is not available under zero data retention.

Where you can use it

At its June 9, 2026 launch, Claude Fable 5 was made generally available on the Claude API, the Claude Platform on AWS, Amazon Bedrock, Google Vertex AI, and Microsoft Foundry — but access was suspended within days (see the update at the top of this guide). When access returns, that spread means you can adopt it inside whichever cloud your data and billing already live in, rather than routing everything through a single endpoint. The capability set is consistent across surfaces: effort, task budgets (beta), the memory tool, code execution, programmatic tool calling, context editing (beta), compaction, and vision are all supported.

There is also a sibling model worth knowing about. Claude Mythos 5 (claude-mythos-5) shares the same capabilities and the same $10/$50 pricing, but omits the safety classifiers, so it does not return the refusal stop reason that Fable 5 can. Mythos 5 is available only in limited release through Anthropic's Project Glasswing, which you can read about on the Project Glasswing page. For most teams, Fable 5 is the model you will actually deploy.

When Fable 5 is worth the premium

Claude Fable 5 earns its price on tasks where reasoning quality directly drives the outcome and a cheaper model would force costly retries or human cleanup. Think multi-hour agentic runs, large-context code refactors, dense research synthesis across hundreds of documents, and any long-horizon plan that must stay coherent from the first step to the last. On work like that, the always-on adaptive thinking and 1M context are the point, and paying double per token is cheaper than the failures you avoid.

For high-volume, low-complexity work, the math flips. Routine classification, short rewrites, and simple extraction rarely benefit from Fable 5's depth, and the premium adds up fast at scale. If you want to map your workloads to the right tier instead of defaulting to the top, see Best Claude Models for OpenClaw and the cost-focused best cheap models guide for 2026. A head-to-head with the next tier down is covered in Claude Fable 5 vs Opus 4.8.

Limitations and Tradeoffs

The clearest tradeoff is price: at $10 input and $50 output per million tokens, Fable 5 costs roughly double Claude Opus 4.8 for the same 1M context and 128K output. Because thinking is always on and cannot be disabled, even moderate tasks carry the reasoning overhead, so the per-task cost is higher than a model where you can turn thinking off. The safety classifiers add another wrinkle: a refusal can interrupt otherwise benign work, and although you are not billed for a pre-output refusal, you still have to handle the fallback path in your code.

There is also a hard compliance limit. Fable 5 requires 30-day data retention as a Covered Model, which rules it out for organizations bound to zero data retention. And for simple, repetitive jobs it is plainly overkill. Do NOT use Fable 5 for high-volume routine classification, short canned responses, bulk simple extraction, or any ZDR-constrained deployment; reach for a cheaper or thinking-optional model in those cases.

Related Guides

• Claude Fable 5 vs Opus 4.8
• Best Claude Models for OpenClaw
• Sakana Fugu vs Claude Fable 5
• The Anthropic Ecosystem Guide

FAQ

How much does Claude Fable 5 cost?

Claude Fable 5 costs $10 per million input tokens and $50 per million output tokens. That is roughly double Claude Opus 4.8, which is $5 input and $25 output for the same 1M context window and 128K output ceiling. Because output is the expensive side, output-heavy workloads feel the premium the most.

Is Claude Fable 5 better than Opus 4.8?

Claude Fable 5 is Anthropic's most capable model and is built to outperform Opus 4.8 on the hardest reasoning and long-horizon agentic tasks. Anthropic has not published specific Fable 5 benchmark numbers at launch, so the difference is best described qualitatively rather than as a single score. For routine work, Opus 4.8 often delivers comparable results at half the price.

What is the Claude Fable 5 context window?

Claude Fable 5 has a 1M-token context window by default. It can also return up to 128K output tokens in a single request. It shares the same tokenizer as Claude Opus 4.8, so token counts are roughly unchanged if you migrate from Opus 4.7 or 4.8.

What is Claude Mythos 5?

Claude Mythos 5 (claude-mythos-5) is a sibling model that shares Fable 5's capabilities and $10/$50 pricing but omits the safety classifiers. It is available only in limited release through Anthropic's Project Glasswing. For nearly all production use, Claude Fable 5 is the model teams deploy instead.

Can I turn off thinking on Claude Fable 5?

No. Adaptive thinking is always on and is the only thinking mode on Claude Fable 5, so thinking:{type:"disabled"} is not supported. You control how deeply it reasons with the effort parameter, which accepts low, medium, high, xhigh, and max. The raw chain of thought is never returned; thinking blocks come back as a summary or empty.