Claude Fable 5: Release Date, Features, Pricing (2026)

Claude Fable 5 is the model Anthropic released on June 9, 2026 for demanding reasoning and long-horizon agentic work — the tier that sits above the Opus line. It is still served, still supported, and still the model a lot of production code points at. But as of September 1, 2026 it is officially a legacy model: Claude Fable 5.1 is the current Fable release, and Anthropic's docs now recommend migrating.

This page is the reference for Fable 5 itself — its release date, specs, pricing, benchmark scores, availability, and the API behaviours that trip up integrations. Every figure below comes from Anthropic's own model pages and pricing table, not from third-party reporting.

What is Claude Fable 5?

Claude Fable 5 is a frontier-tier Anthropic model built for demanding reasoning and long-horizon agentic work — tasks that run for hours with minimal human intervention. It sits above the Opus family rather than inside it, which is why it carries its own name instead of an Opus version number.

There are two models in the family:

• Claude Fable 5 (claude-fable-5) — the public model, available to every customer without access approval.
• Claude Mythos 5 (claude-mythos-5) — the same underlying model without Fable 5's safety classifiers, offered by invitation only through Project Glasswing for defensive cybersecurity workflows.

The distinction matters and is often reported wrong: Mythos 5 is not a more capable model. It shares Fable 5's specs, pricing, and capabilities exactly. The only difference is that Fable 5 ships with classifiers that can decline requests and Mythos 5 does not. Our Anthropic Mythos guide covers that programme in detail.

When was Claude Fable 5 released?

Claude Fable 5 was released on June 9, 2026, alongside Claude Mythos 5. It went live the same day on the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry.

Two later dates matter as much as the launch date:

• June 12, 2026 — Anthropic pulled Fable 5 and Mythos 5 offline entirely (see below).
• July 1, 2026 — access was restored, and has been continuous since.
• September 1, 2026 — Claude Fable 5.1 shipped, and Fable 5 moved to legacy status.

Is Claude Fable 5 still available, and when does it retire?

Yes. Fable 5 is listed as Active (legacy) in Anthropic's model deprecations table. It is not deprecated, has no announced deprecation date, and requests to claude-fable-5 work normally on all five platforms.

Its retirement is committed to "not sooner than June 9, 2027" — one year from launch. In Anthropic's lifecycle vocabulary, "legacy" means the model will no longer receive updates and may be deprecated in future; "deprecated" is the next stage and comes with a firm retirement date and at least 60 days' notice by email. Fable 5 has not reached that stage.

Practical read: there is no urgency to rip claude-fable-5 out of a working integration this month. There is a good reason to schedule the move to 5.1 — the cache-read price — but nothing is about to break.

What are Claude Fable 5's features and specs?

The 1M-token window is billed at standard per-token rates across its whole length — there is no long-context premium. Note the tokenizer: Fable 5 uses the one introduced with Claude Opus 4.7, which produces roughly 30% more tokens for the same text than pre-4.7 models. If you sized max_tokens or a context budget against an older model, re-measure rather than assuming.

Supported features include the effort parameter, task budgets (beta), the memory tool, code execution, programmatic tool calling, context editing (beta), compaction, and vision.

How much does Claude Fable 5 cost?

Fable 5 is priced at exactly double Claude Opus 5 and Claude Opus 4.8 on both input and output.

The cache-read row is the one worth staring at. Every Claude model prices cache hits at 0.1× base input — except Fable 5.1 and Mythos 5.1, which price them at 0.025×. On Fable 5 a cache hit costs $1.00 per million tokens; on Fable 5.1 the identical hit costs $0.25. Nothing else in the price list changes between the two versions.

That has a specific shape to it. If your workload is single-shot — one prompt, one answer, no reuse — the two versions cost the same. If it is a long agentic session that re-reads a large cached prefix on every turn, which is exactly what Fable is built for, the cache-read line dominates your bill and 5.1 quarters it. Anthropic's own estimate is ~25% total savings for typical workloads and ~45% for heavily agentic ones.

Batch API requests take 50% off input and output on both versions. Specifying US-only inference via inference_geo applies a 1.1× multiplier across every pricing category. Full rates are on Anthropic's pricing page. For subscription-plan limits rather than API rates, see our breakdown of Claude Fable 5 usage credits and plan limits.

How does Claude Fable 5 score on benchmarks?

Anthropic published a directly comparable set of numbers with the Fable 5.1 announcement, which is the cleanest apples-to-apples source available — the same evals run across Fable 5.1, Fable 5, Claude Opus 5, and GPT-5.6 Sol.

Source: Anthropic's Claude Fable 5.1 and Mythos 5.1 announcement. Claude Mythos 5.1 scores 60.9% on Terminal-Bench 4.0, the highest figure Anthropic published.

Two things stand out, and neither flatters Fable 5. First, the jump from Fable 5 to 5.1 is large on agentic work specifically — Terminal-Bench 4.0 goes 42.0% → 55.8%, AutomationBench nearly doubles, and Terminal-Bench-Science more than doubles. Second, and more awkward for anyone still on Fable 5: Claude Opus 5 beats Fable 5 on most of these evals while costing half as much. Opus 5 leads on Terminal-Bench 4.0, GDPval-AA v2, AutomationBench, OSWorld, and Terminal-Bench-Science; Fable 5 holds narrow leads only on CursorBench (70.5% vs 70.0%) and Humanity's Last Exam (57.8% vs 56.6% untooled).

Fable 5 was released six weeks before Opus 5, which explains it. But it means the honest position today is that Fable 5 is not the "buy the expensive one for better results" option it was in June. We work through that trade in full in Claude Opus 5 vs Claude Fable 5.

Why did Claude Fable 5 go offline in June 2026?

Three days after launch, on June 12, 2026, Anthropic pulled both Fable 5 and Mythos 5 offline for all users. The reason was not a capability failure: the US government applied export controls to the two models requiring that access be restricted from foreign nationals, and Anthropic had no reliable way to verify nationality in real time. Rather than partially enforce it, the company suspended access entirely.

Access was restored on July 1, 2026. On redeployment Anthropic shipped an improved safety classifier targeting a specific jailbreak technique that had been reported by Amazon researchers, which it says blocked the reported behaviour in over 99% of cases. It also announced deeper pre-release testing with US government partners and an industry framework for grading jailbreak severity, developed with Amazon, Microsoft, Google and others. Details are in Anthropic's redeployment statement; we covered the outage as it resolved in Is Claude Fable 5 back?

The episode is worth knowing about for one practical reason: it is the clearest illustration to date that frontier-tier model access can be withdrawn on regulatory grounds with no notice. If Fable 5 or 5.1 is a single point of failure in your product, configure a fallback.

What changes in your code when you call Claude Fable 5?

Fable 5 is not a drop-in swap for an Opus-tier model. Five behaviours differ, and four of them return a 400 if you get them wrong.

• Thinking is always on. Adaptive thinking cannot be turned off. Both thinking: {"type": "disabled"} and {"type": "enabled", "budget_tokens": N} return a 400. Omit the parameter or send {"type": "adaptive"}, and control depth with output_config.effort instead. Default effort is high.
• Raw reasoning is never returned. thinking.display defaults to "omitted", which returns thinking blocks with an empty thinking field — to a streaming UI that looks like a long silent pause before output. Set "summarized" to get a readable summary. The raw chain of thought is never exposed at any setting.
• Assistant prefill returns a 400. Use structured outputs (output_config.format) or a system-prompt instruction instead.
• Sampling parameters return a 400. Non-default temperature, top_p, or top_k are rejected. Steer with prompting.
• 30-day data retention is mandatory. Fable 5 is not available under zero data retention unless Anthropic expressly authorises it. An organisation configured for ZDR gets a 400 invalid_request_error on every Fable 5 request, with a perfectly valid payload — check retention configuration before debugging the request body.

Handling refusals

Fable 5's safety classifiers can decline a request. When they do, the Messages API returns a successful HTTP 200 with stop_reason: "refusal" and a stop_details object naming the policy area that fired. Code that reads response.content[0] unconditionally will break — check stop_reason first.

Three classifier categories exist: cybersecurity exploitation, biology and chemistry, and model distillation. Anthropic says they trigger in under 5% of sessions on average. You are not billed for a refusal that arrives before any output.

The fix is the fallbacks parameter, which retries a declined request on another model server-side inside the same call. fallbacks: "default" routes by refusal category to Anthropic's recommended model. Fallback credit refunds the prompt-cache cost of the switch so you don't pay it twice. Configure this by default rather than as later hardening — benign security and life-sciences work does occasionally trip the classifiers.

What's different in Claude Fable 5.1?

Fable 5.1 keeps the specs, the tokenizer, the price, and the API surface almost intact. The migration is a model-ID change plus three checks.

Three breaking changes:

• Forced tool use is gone. tool_choice of {"type": "any"} or {"type": "tool", "name": "..."} returns a 400. Use {"type": "auto"} with strict: true, or move the schema to structured outputs, and name the tool in the prompt.
• Thinking blocks are bound to the model that made them. Fable 5.1 can read earlier models' thinking; no earlier model can read Fable 5.1's. A conversation that moves onto 5.1 keeps its reasoning; one that moves off it loses reasoning for the turns that run elsewhere.
• Editing earlier turns invalidates thinking blocks. Rebuilding system or tools mid-conversation, or injecting-then-removing per-request text, breaks every later thinking block. Treat the conversation as append-only. This check is enforced for accounts created on or after August 31, 2026.

Five additive changes: the cache-read price cut, per-message effort (beta), turn-scoped system messages (beta), progress updates between tool calls via display: "updates" (beta), and content provenance — a statistical text watermark on all output plus C2PA Content Credentials on generated media.

Behaviour also shifts in ways no code change surfaces: 5.1 batches parallel tool calls less reliably, writes fewer progress updates, answers from memory more often at low effort, and is likelier to rewrite a whole file than make a targeted edit. Each has a prompting fix in Anthropic's what's-new page.

Should you use Claude Fable 5, Fable 5.1, or Opus 5?

A decision rule, in the order Anthropic itself now recommends:

• Start with Claude Opus 5. Anthropic's own guidance is that most workloads should start here. It is half Fable's price and beats Fable 5 on most published evals. Tune effort before you consider switching models — that is usually the better lever.
• Move to Claude Fable 5.1 when your evals on Opus 5 at xhigh or max effort still fall short on demanding reasoning or long-horizon agentic work. Agent sessions that run for hours, multistep deep research, and analysis carried through to a finished document or deck are the named use cases.
• Stay on Claude Fable 5 only if you have a specific reason — a pinned eval baseline, a compliance review tied to that snapshot, or an integration whose history-editing pattern would break 5.1's thinking-block binding and that you can't fix yet. Otherwise 5.1 is strictly better at the same headline price.

The one thing not to do is treat Fable 5 as the default premium choice. That was true in June 2026. Two model releases later it isn't: for the same money you can run Opus 5 twice, and for the same money as Fable 5 you can run Fable 5.1 with cache reads at a quarter of the cost.

FAQ

When was Claude Fable 5 released?

June 9, 2026, alongside Claude Mythos 5. It was taken offline on June 12, 2026 following US export controls and restored on July 1, 2026. Claude Fable 5.1 succeeded it on September 1, 2026, at which point Fable 5 became a legacy model.

How much does Claude Fable 5 cost?

$10 per million input tokens and $50 per million output tokens — exactly double Claude Opus 5. Cache writes are $12.50 per million for the 5-minute TTL and $20 for the 1-hour; cache reads are $1.00 per million. Batch API requests take 50% off both input and output.

Is Claude Fable 5 still available?

Yes. It is listed as Active (legacy) and works normally on the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry. It has not been deprecated, and Anthropic commits to retiring it no sooner than June 9, 2027.

What is Claude Fable 5's context window?

1,000,000 tokens, which is both the default and the maximum, billed at standard per-token rates across the full window with no long-context premium. Maximum output is 128,000 tokens per request. The minimum cacheable prompt length is 512 tokens.

What is the difference between Claude Fable 5 and Claude Mythos 5?

They are the same underlying model with identical specs, pricing, and capabilities. Fable 5 ships with safety classifiers that can decline requests; Mythos 5 does not. Mythos 5 is invitation-only through Project Glasswing for defensive cybersecurity workflows, while Fable 5 needs no access approval.

Why do some Claude Fable 5 requests come back refused?

Fable 5's classifiers cover cybersecurity exploitation, biology and chemistry, and model distillation, and trigger in under 5% of sessions. A declined request returns HTTP 200 with stop_reason: "refusal", not an error, and is not billed if it arrives before any output. Use the fallbacks parameter to retry on another model automatically.

Can I use Claude Fable 5 with zero data retention?

No. Fable 5 requires 30-day data retention and is not available under zero data retention unless Anthropic expressly authorises it. An organisation configured for ZDR receives a 400 error on every Fable 5 request regardless of payload — check the retention setting before debugging anything else.

Should I migrate from Claude Fable 5 to Claude Fable 5.1?

For most workloads, yes. The headline price is identical but cache reads cost a quarter as much, and 5.1 scores higher on every published benchmark. Budget time for three breaking changes: forced tool use now errors, thinking blocks are bound to the model that produced them, and editing earlier turns invalidates them.

Comparing against OpenAI? See our GPT-5.6 vs Claude Fable 5 breakdown for pricing, benchmarks, and where each model wins.