Claude Fable 5

Claude Fable 5 is Anthropic's highest-intelligence widely available model, built for long-running agents. Specs, $10/$50 pricing and refusal behaviour.

Claude Fable 5 is Anthropic's highest-intelligence widely available model, generally available since 9 June 2026. It costs $10 per million input tokens and $50 per million output tokens — exactly twice Claude Opus 5 — with a 1,000,000-token context, 128,000-token maximum output, a January 2026 knowledge cutoff and adaptive thinking that is always on and cannot be turned off.

It is built for one thing: long-running agents that work autonomously for hours. It is also the slowest model Anthropic sells, at roughly 75 output tokens per second, with materially tighter rate limits than Opus 5. Most workloads should not use it.

#What "long-horizon agent model" actually means

The phrase is doing real work, not marketing. A long-horizon agent runs unsupervised across many steps — planning, calling tools, checking its own output, recovering from failures — for hours without a human turn. The failure mode that matters is not a wrong sentence; it is a wrong decision at step 40 that quietly corrupts the next 200.

Fable 5 is tuned for that shape of work rather than for conversation. Anthropic positions it for deep reasoning, long-horizon agentic tasks and advanced research, and it supports the features those workloads need: programmatic tool calling, the memory tool, and task budgets (beta header task-budgets-2026-03-13) in place of the injected context-remaining tags Sonnet-class models receive.

#Why it costs twice as much as Opus 5

Per token, the arithmetic never favours Fable 5. Per completed task it sometimes does, under one condition: when the cost of a failed run exceeds the cost of the tokens. A four-hour run that produces a wrong migration costs the tokens twice plus the time to diagnose it. If a better model cuts that failure rate enough, the 2× premium disappears.

That condition holds less often than teams assume. Anthropic itself states that Opus 5 at maximum effort lands within 0.5% of Fable 5's peak CursorBench 3.2 score at half the cost per task, and surpasses Fable 5's best OSWorld 2.0 result at just over a third of the cost. The vendor is telling you the cheaper model usually wins on value. The model comparison page works that framing through in detail.

Note the last row: the output-300k-2026-03-24 beta header that raises Batch max_tokens to 300,000 covers Opus 5, Opus 4.8, 4.7, 4.6 and Sonnet 5 — but not Fable 5, which stays at 128,000. The most expensive model has the smaller extended-output ceiling. Caching and batching mechanics are on the Claude API page.

#Always-on adaptive thinking

On Fable 5, thinking: {"type": "disabled"} is not supported. Adaptive thinking is permanently on, and the only control you have is the effort ladder — low, medium, high, xhigh, max. Opus 5 by contrast allows thinking to be switched off at effort high or below, and Haiku 4.5 still uses manual extended thinking with an explicit token budget.

Two consequences follow. First, thinking tokens are billed as output tokens at $50 per million and count against max_tokens and your rate limits — on Fable 5 you always pay for reasoning, whether the task needed it or not. Second, thinking.display defaults to "omitted", and the raw chain of thought is never returned on Fable 5 or Mythos 5 under any setting — you get an encrypted signature instead. If your compliance posture requires inspectable reasoning traces, this model cannot provide them.

#Fable 5 refusals, and the fallbacks parameter

Fable 5 handles declined requests differently from every model before it, and this is the single change most likely to break an existing integration.

The billing treatment is fair: you are not billed for a request refused before any output was generated. And Anthropic provides three retry paths. Server-side fallback uses a fallbacks parameter behind the server-side-fallback-2026-07-01 beta header, including a "default" mode that follows an Anthropic-recommended chain of models. Client-side fallback runs in SDK middleware. Manual fallback is you, catching the stop reason and re-issuing the call.

When a fallback fires, a "fallback credit" refunds the prompt-cache cost of switching — the cache warmed on Fable 5 is not reusable on the fallback target, so Anthropic credits that write rather than charging twice for the same context. For an agent loop with a large cached system prompt, that is not a rounding error.

Fable 5 is not alone in this. Sonnet 5 is the first Sonnet-tier model with real-time cybersecurity safeguards and can also return stop_reason: "refusal", so handle the stop reason once, for every model, rather than as a Fable-specific quirk. Retry and orchestration patterns are on the automation page.

#Rate limits: much tighter than Opus 5

This is the practical ceiling on Fable 5 adoption, and it is easy to miss when planning capacity. Fable 5 sits in its own rate-limit bucket with limits well below every other current model at every tier.

On the Start tier that is 500,000 input tokens per minute against Opus 5's 2,000,000, and a quarter of the output allowance. Combine that with ~75 tokens per second of generation and Fable 5 is unsuitable for anything interactive or bursty. Cache reads do not count toward the input-tokens-per-minute figure on any current model, so a well-cached agent gets more real throughput than the raw number suggests — but the gap remains at every tier. Tier names and spend caps are on the pricing page.

#What is Claude Mythos 5?

Claude Mythos 5 (claude-mythos-5) is the same underlying model as Fable 5 with certain safety classifiers removed. Fable 5 ships with classifiers that can decline requests; Mythos 5 does not include them. Anthropic offers it, and the earlier Claude Mythos Preview, for defensive cybersecurity workflows under a programme called Project Glasswing.

The facts worth stating plainly:

• Specifications and pricing are identical to Fable 5 — 1M context, 128k output, $10/$50.
• Access is invitation-only, with no self-serve sign-up. It is limited to approved organisations.
• Both Fable 5 and Mythos 5 are Covered Models: 30-day data retention, and not available under zero data retention. If your contract depends on ZDR, neither model is an option.
• Anthropic states export controls on Mythos 5 were lifted on 1 July 2026, restoring access for approved US organisations.

That is the whole picture as documented. Treating a model with fewer refusal classifiers as a general-purpose upgrade misreads what it is for: it exists because defensive security research collides with safety classifiers constantly, not because unfiltered output is better. Organisations with requirements in this territory should talk to Anthropic through the enterprise route.

#Should you use Fable 5?

Most people should not. Fable 5 suits a narrow band of work: long-running autonomous agents where one failure costs more than the entire token bill, deep research runs, and the reasoning tasks where it demonstrably beats Opus 5. Outside that band, Opus 5 gets you close for half the price with four times the rate-limit headroom and newer knowledge, and Sonnet 5 handles most production traffic at a fifth of the cost.

The case for it is independent: it holds first place on arena.ai's Elo leaderboard at 1507, and Anthropic cites Cognition's FrontierCode and the Hebbia Finance Benchmark as places where it takes the highest score of any frontier model. The case against it is equally documented — on Agents' Last Exam the leading OpenAI model scores 52.7% to Fable 5's 40.5%, and Google's Gemini 3.7 Flash generates roughly 3,900 tokens per second against Fable 5's 75.

A sensible pattern is to use it surgically: let Sonnet 5 or Opus 5 drive the agent loop and call Fable 5 only for the planning or review step that justifies the cost, keeping the rate-limit ceiling out of your critical path. Orchestration patterns are covered under MCP agents and Claude for developers. Note also that Fable 5 is excluded from the free plan entirely. The rest of the lineup is indexed on the Claude models page.

#Frequently asked questions