Claude Fable 5: Anthropic's First Mythos-Class Model

Claude Fable 5 is Anthropic's first publicly available Mythos-class model, released June 9, 2026 alongside its no-classifier sibling Claude Mythos 5. It is the first model in the Claude 5 family and sits a tier above Anthropic's Opus class in capability. It matters because independent benchmarks rank it first overall, with coding scores that clear every prior Claude model by a wide margin.

What is Claude Fable 5?

Claude Fable 5 is the first publicly available model in what Anthropic calls its Mythos-class tier, a level that sits above the Opus class in raw capability. Anthropic describes Fable 5 as a Mythos-class model made safe for general use, which is the practical distinction from its locked-down sibling. It is also the first model in the Claude 5 family, so it sets the baseline for everything that follows in this generation.

Anthropic announced Claude Fable 5 and Claude Mythos 5 together on June 9, 2026. Fable 5 is the one you can actually reach: it is generally available across the Claude API and major cloud platforms. The headline pitch is that a Mythos-class model is now packaged for everyday developer and business use, with safety guardrails layered on top rather than a separate, weaker model standing in for it.

If you want the primary source, the full launch details live in Anthropic's official announcement and the Claude model documentation.

Fable 5 vs Mythos 5: what's the difference?

Fable 5 and Mythos 5 are the same underlying model. There is no quality gap between them. The only difference is safety: Claude Fable 5 includes classifier-based safeguards that can decline certain requests, and Claude Mythos 5 ships without those classifiers. Same weights, same reasoning, different guardrails.

That design choice drives who gets each model. Fable 5 is generally available. Mythos 5 is a limited release, distributed only through a program Anthropic calls Project Glasswing, the successor to its earlier Claude Mythos Preview. Project Glasswing puts the unclassified model in the hands of approved cyberdefenders and critical-infrastructure providers, with a planned expansion to biology researchers.

For most builders the takeaway is simple: you will be using Claude Fable 5, and you get the full Mythos-class model with a thin safety layer on top. The two share the same context window, the same pricing, and the same adaptive reasoning. The classifiers are the entire delta, and as you will see below, they trigger rarely.

How good is Claude Fable 5? The Fable 5 benchmarks

Third-party testing is where Fable 5 separates from the pack. The independent Artificial Analysis Intelligence Index scored it 64.9 and ranked it #1 overall. On Humanity's Last Exam it reached 53%. The Fable 5 benchmarks that will matter most to engineers, though, are the coding ones, where the gap over the previous generation is largest.

On SWE-bench Pro, Fable 5 vs Opus 4.8 is an 11-point swing in Fable 5's favor, and it leads GPT-5.5 and Gemini 3.1 Pro by more than 20 points. On Cognition's FrontierCode it more than doubles Opus 4.8. On the Every Senior Engineer benchmark it scores 91 against Opus 4.8's 63. These are different harnesses measuring different things, but they point the same direction.

Anthropic backs the numbers with partner anecdotes, which we report as Anthropic and partner claims rather than as independently verified facts. Anthropic says Stripe used the model to complete a 50-million-line Ruby migration in about a day against an estimated two months, and frames the model as compressing months of engineering into days. Anthropic also reports it played Pokemon FireRed using vision alone, where older Claude models needed helper harnesses, and that on Slay the Spire with file-based memory it roughly tripled Opus 4.8's performance, reaching the final act three times as often. On drug design, Anthropic says internal experts reported accelerating parts of the process by around 10 times.

How much does Claude Fable 5 cost?

Fable 5 pricing is $10 per million input tokens and $50 per million output tokens, and the same rates apply to Mythos 5. That is roughly 2x the price of Claude Opus 4.8. Anthropic frames it differently, pointing out the model is less than half the price of Claude Mythos Preview. Whether it reads as expensive or cheap depends on what you are comparing it to.

On subscriptions, the rollout came with a free window. From June 9 to June 22, 2026, the new models were included at no extra cost on Pro, Max, Team, and seat-based Enterprise plans. From June 23 they require usage credits, with broader restoration promised when capacity allows. On consumer Claude.ai plans, the new models count as 2x usage against your limit.

Where can you use Fable 5, and what are the specs?

The API model IDs are claude-fable-5 and claude-mythos-5. Fable 5 is generally available through the Claude API, Claude on AWS via Amazon Bedrock, Vertex AI on Google Cloud, and Microsoft Foundry. It is also reported to be available in Claude Code, which we note softly as a report rather than a confirmed channel.

On capabilities, the context window is 1,000,000 tokens, with a maximum output of 128,000 tokens per request. Adaptive thinking is always on, and you set its depth through an effort parameter rather than toggling reasoning off. The raw chain-of-thought is never returned, so you get the answer and the effort dial but not the internal trace.

On data handling, both models use 30-day retention and are designated Covered Models. That designation matters for one practical reason: they are not available under zero-data-retention agreements, so teams that require ZDR will need to plan around that constraint before building on them.

What are Claude Fable 5's safety guardrails?

The classifiers that define Fable 5 cover three domains: cybersecurity, biology and chemistry, and distillation, the last of which is about preventing capability extraction to train rival models. These are the safeguards that Mythos 5 omits, and they are the reason Fable 5 is described as made safe for general use.

The fallback behavior is the part worth understanding. When a safeguard triggers, the request does not just error out - it falls back to Claude Opus 4.8 to complete the response. Anthropic says the safeguards trigger on average in less than 5% of sessions and that they are tuned conservatively, which means they sometimes catch harmless requests. For most general workloads you will rarely see them, but if your prompts touch security tooling or chemistry you may occasionally land on the Opus 4.8 path instead of the full Mythos-class model.

What happened to Claude Fable 5 after launch?

Three days after launch, on June 12, 2026, a US government export-control directive forced Anthropic to suspend access to both Claude Fable 5 and Claude Mythos 5. The suspension landed in the middle of the free subscription window and changed the picture for anyone who had started building on the model. We cover the directive, what it means for existing integrations, and where access stands now in our companion post: Claude Fable 5 and Mythos 5 access suspended. Anthropic's own statement is available in its access update.