Claude Fable 5: Anthropic’s Most Powerful Public Model, Explained

Claude Fable 5 is the most capable model Anthropic has ever made generally available, launched on June 9, 2026 as the safe-for-everyone twin of the restricted Claude Mythos 5. According to Anthropic’s launch announcement, Fable 5 is “state-of-the-art on nearly all tested benchmarks,” excelling at software engineering, knowledge work, vision, and scientific research.

In one line: Fable 5 is a Mythos-class frontier model priced at $10 per million input tokens and $50 per million output tokens, with a 1M-token context window. It is built for multi-day autonomous coding and knowledge work rather than everyday chat, which makes it a very different tool from the faster, cheaper Claude Sonnet 5 that most people use day to day.

What Is Claude Fable 5?

Fable 5 is Anthropic’s frontier model for the hardest coding and knowledge-work problems. It shares its underlying weights with Claude Mythos 5, a restricted release available only to vetted cybersecurity and biology partners. The two models are the same brain; Fable 5 is simply the version wrapped in safety safeguards so it can be released to the public.

The Mythos-class twin

Anthropic describes Fable 5 as a model whose capabilities “exceed those of any model” it has previously made generally available. That framing is deliberate: Mythos-class weights were powerful enough that Anthropic paired the public release with active monitoring rather than shipping the raw model. Fable 5 and Mythos 5 launched together on June 9, 2026, with Fable 5 available immediately on the Claude API.

The practical difference is not raw intelligence but access. Mythos 5 removes the general safeguards for approved researchers who need unfiltered output for defensive security and biology work. Fable 5 keeps those safeguards on for everyone else.

Where it sits in the Claude 5 family

Fable 5 is one member of Anthropic’s newest lineup, alongside Sonnet 5 (model id claude-sonnet-5, released June 30, 2026), the flagship Opus 4.8, and the lightweight Haiku 4.5. Each tier answers a different need, and Fable 5 sits at the top for capability and price.

If Sonnet 5 is the model you reach for to run everyday agents cheaply, Fable 5 is the one you reserve for problems that would otherwise take a skilled person days.

Benchmarks and Capabilities

Anthropic and its evaluation partners position Fable 5 at the top of the field across coding, analytics, vision, and computer use. The headline claim — state-of-the-art on nearly every tested benchmark — is backed by several independent evaluations.

State-of-the-art across coding, knowledge work, and vision

Fable 5 posted the highest score on Cognition’s FrontierBench coding evaluation, and it became the first model to break 90% on Hex’s core analytics benchmark for complex tasks. It also ranked as the strongest model on Hebbia’s Finance Benchmark for senior-level reasoning. On the widely watched SWE-Bench Pro software-engineering test, secondary reporting puts Fable 5 in the low-80s percent range versus roughly the high-50s for OpenAI’s GPT-5.5, though the precise figures should be treated as approximate until Anthropic publishes the full card.

The through-line across these results is that Fable 5 does best on long, messy, multi-step problems rather than trivia. It reads diagrams, charts, and tables inside files and PDFs, and it can drive a computer through a vision-only interface.

Built for long, autonomous work

The defining feature of Fable 5 is endurance. Anthropic says the model can “work for days at a time: planning across stages, delegating to sub-agents, and checking its own work.” It is tuned for long-running, complex, asynchronous tasks that previously required frequent human check-ins.

That self-checking behavior matters as much as the raw score. A model that validates its own output before returning it wastes fewer of your expensive output tokens on errors you would otherwise have to catch and re-prompt.

Fable 5’s capabilities exceed those of any model we have previously made generally available.

Anthropic

Typical jobs that fit this profile include large code migrations, multi-day autonomous coding sessions, deep research, and document-heavy finance or legal work where the context runs into hundreds of thousands of tokens.

Pricing and Context Window

Fable 5 is priced as a premium frontier model, not an everyday-agent model. The gap between it and the rest of the Claude 5 family is the single most important thing to understand before you send it a request.

What Fable 5 costs

Fable 5 costs $10 per million input tokens and $50 per million output tokens — roughly double Opus 4.8’s $5/$25 and far above Sonnet 5’s introductory $2/$10. Anthropic applies a 90% prompt-caching discount on cached input, which softens the cost for workloads that reuse a large stable context. US-only inference is available at 1.1x the standard price for teams with data-residency requirements. The exact numbers are published on the Claude Platform pricing page.

The takeaway is simple:

• Use Sonnet 5 or Haiku 4.5 for high-volume, routine agent traffic.
• Use Opus 4.8 for demanding work at a middle price point.
• Reserve Fable 5 for the few problems where its frontier edge pays back the premium.

Context and outputs

Fable 5 ships with a 1,000,000-token context window and supports up to 128k output tokens in a single request. Inputs can be text, images, or files, and the model returns text with reasoning support. That million-token window is what enables the multi-day sessions: an entire codebase, a stack of PDFs, or a long research corpus can sit in context at once.

Combined with the 90% caching discount, the large window changes the economics of long tasks. If the bulk of your context is stable across many calls, most of your input cost collapses to a tenth of the sticker price.

Availability and the Rollout Pause

Fable 5 launched broadly but hit demand-driven turbulence within days. Knowing where it runs — and why access wobbled — helps set expectations.

Where you can use it

Fable 5 is available on the Claude API, the Claude Platform on AWS, Amazon Bedrock, Google Cloud, and Microsoft Foundry. That multi-cloud spread means most enterprises can call it through the provider they already use. As Anthropic detailed at launch, through roughly June 22, 2026 it was also included in Pro, Max, Team, and seat-based Enterprise plans at no extra cost; from June 23, subscription access shifted to usage credits.

Why demand paused access

Interest in a model that can run autonomously for days quickly outstripped Anthropic’s capacity. Shortly after launch the company throttled and then temporarily paused access to Fable 5, stating it was “working to restore access as soon as possible” and apologizing for the disruption. The pause was an infrastructure decision, not a product recall — the model itself was unchanged, and Anthropic signaled it intended to restore Fable 5 as a standard feature.

Safety Safeguards

Because Fable 5 shares Mythos-class weights, Anthropic wrapped it in an unusually active safety layer. Rather than blanket refusals, the system quietly reroutes the small slice of high-risk traffic to a less capable model.

How the fallback works

A classifier watches every session for high-risk requests in three areas: offensive cybersecurity, biology and chemistry, and distillation attacks that try to copy the model. When it flags a request, the session falls back to Claude Opus 4.8 rather than refusing outright. Anthropic reports these safeguards trigger in less than 5% of sessions on average, meaning at least 95% of Fable sessions run entirely on Fable 5.

Here is what actually happens in practice:

1. You send a request to Fable 5.
2. A classifier scores the session for cybersecurity, bio/chem, and distillation risk.
3. If the score stays low, Fable 5 answers directly.
4. If the score is high, the response comes from Opus 4.8 instead.
5. You keep getting a useful answer, just from a more conservative model.

Red-teaming and data retention

Before launch, Fable 5 went through more than 1,000 hours of external bug-bounty testing, and outside red-teaming organizations reported finding no universal jailbreaks. Anthropic also applies a 30-day traffic-retention policy to Mythos-class models as a defense against novel attacks; that retained data is used for security monitoring, not for training. For teams in regulated industries, that retention window is worth confirming against internal policy before adopting Fable 5 at scale.

FAQ