Fable 5 / Mythos 5

Fable, you say? But where’s Mythos? Fable 5 is the first Mythos-class model the public has ever been allowed to touch; Mythos 5 is the same model with the muzzle off, and reserved for vetted defenders. Together they’re the most capable AI models in existence (and the first one that comes with a designated stunt double).

What shipped

Models: Claude Fable 5 (claude-fable-5 on the Claude API, anthropic.claude-fable-5 on Bedrock, claude-fable-5 on Vertex AI) and Claude Mythos 5 (claude-mythos-5). Same underlying model: Fable 5 ships with classifier safeguards, Mythos 5 ships with some of them lifted and succeeds the invitation-only Claude Mythos Preview (claude-mythos-preview).

Model type: Text + image input, text output, for both models. Adaptive thinking is always on.

Pricing: $10 / $50 per million input / output tokens on the Claude API, for both Fable 5 and Mythos 5. That’s double Claude Opus 4.8’s $5 / $25. Fable 5 is included at no extra cost on Pro, Max, Team, and seat-based Enterprise plans through June 22; after that, subscribers buy extra usage credits to keep access.

Available on: The Claude app (Pro, Max, Team, and Enterprise plans, phased rollout through June 22), the Claude API, Claude Platform on AWS, Amazon Bedrock, Google Vertex AI, and Microsoft Foundry. Mythos 5 is limited-availability only, through Project Glasswing.

Other info: 1M-token context window, 128K max output, for both models. Uses the tokenizer introduced with Opus 4.7, so the same text produces roughly 30% more tokens than pre-4.7 models. Fable 5’s classifier safeguards cover three domains (cybersecurity, biology/chemistry, and distillation); when they trigger, the response is silently handled by Claude Opus 4.8 instead. Anthropic says the safeguards fire in under 5% of sessions. System card published covering both models. No ASL designation stated in the launch announcement. Knowledge cutoff not disclosed.

Anthropic released Claude Fable 5 and Claude Mythos 5 together today, and with them the public’s first access to the Mythos family, the model class that spent the spring spooking the cybersecurity world with its ability to find and chain exploits in minutes. I covered the original Mythos reveal back in March, when Anthropic held the full model back and released only a restricted preview to Project Glasswing partners for defensive security work. Fable 5 is that model made safe for general use: same underlying weights, wrapped in classifiers that detect requests touching cybersecurity, biology and chemistry, or distillation, and quietly hand those requests to Claude Opus 4.8 instead.

The evidence Anthropic put forward is heavy on third-party workload proofs. SWE-Bench Pro at 80.3% puts more than 20 points between Fable 5 and GPT-5.5. Stripe compressed a two-month, 50-million-line Ruby migration into a single day. Hex reports the first score above 90% on its analytics benchmark, and Hebbia gave it the top mark on its finance reasoning suite. On the restricted side, Mythos 5 produced nine strong protein design candidates out of 14 in a drug design evaluation.

But: the price doubles Opus 4.8, subscription access is free only until June 22 before flipping to a credits model, the fallback safeguard means you don’t always know which model answered you, and every Mythos-class prompt now sits in a 30-day retention window (whether you like it or not).

What’s new

This isn’t an Opus increment. It’s a different model family arriving in public for the first time, and the launch architecture around it is as novel as the model.

• One model, two releases. Until today the Mythos family existed behind an invitation-only program for cyberdefenders and critical infrastructure operators. Now it ships in two trims at once: Fable 5 for anyone with a Pro subscription, Mythos 5 for vetted defenders with the cyber restrictions lifted. Anthropic says the capabilities exceed anything they’ve ever made generally available, and the tiered-trust release structure is itself a first for a frontier lab.

• The two-model sandwich. When Fable 5’s classifiers flag a request, Claude Opus 4.8 answers instead. No frontier lab has shipped a designated fallback model as a launch feature before. It’s a genuinely new deployment pattern, and it means “which model am I talking to” now has a probabilistic answer.

• A distillation classifier. One of the three safeguard domains exists specifically to stop competitors from extracting Fable 5’s capabilities into their own models. Capability protection as a safety category is new, and it says a lot about where Anthropic thinks the competitive threat lives.

• Mandatory 30-day retention. All Mythos-class traffic is retained for 30 days for safety monitoring, with automatic deletion after. Anthropic says it’s not used for training. It’s still a policy shift that lands on every user of the model, including the ones who never trip a classifier.

• A price reset at the top. $10 / $50 per million tokens doubles the Opus 4.8 frontier price. The last several frontier releases held the line at $5 / $25. This one moves it.

The benchmarks

Anthropic’s numbers

Anthropic published one column for both models, showing the higher of the two, and says they land within 1 to 3 points of each other. The rows below are the ones it flags as splitting further, because Fable 5 hands cyber and biology requests to Opus 4.8, so they are Mythos 5’s numbers. The rest of the shared column: SWE-Bench Pro 80.3%, FrontierCode (Diamond) 29.3% at xhigh, GDPval-AA 1932, GDP.pdf 29.8%, Blueprint-Bench 2 38.6%, AutomationBench 17.4%, OSWorld-Verified 85.0% and the Legal Agent Benchmark 13.3%.

• Humanity's Last Exam (no tools, Mythos 5): 59.0%. +2.2 over Claude Mythos Preview (56.8%), +9.2 over Claude Opus 4.8 (49.8%), +17.6 over GPT-5.5 (41.4%), +14.6 over Gemini 3.1 Pro (44.4%).

• Humanity's Last Exam (with tools, Mythos 5): 64.5%. −0.2 under Claude Mythos Preview (64.7%), +6.6 over Claude Opus 4.8 (57.9%), +12.3 over GPT-5.5 (52.2%), +13.1 over Gemini 3.1 Pro (51.4%).

• Terminal-Bench 2.1 (Mythos 5): 88.0%. +5.3 over Claude Opus 4.8 (82.7%), +4.6 over GPT-5.5 in Codex CLI (83.4%), +17.3 over Gemini 3.1 Pro in Gemini CLI (70.7%).

• ExploitBench (Mythos 5, capability %): 78.0%. +9.0 over Claude Mythos Preview (69.0%), +38.0 over Claude Opus 4.8 (40.0%), +44.0 over GPT-5.5 (34.0%).

• BioMysteryBench hard (Mythos 5): 46.1%. +16.5 over Claude Mythos Preview (29.6%), +6.1 over Claude Opus 4.8 (40.0%).

• BioMysteryBench human-solved (Mythos 5): 83.9%. +1.3 over Claude Mythos Preview (82.6%), +3.5 over Claude Opus 4.8 (80.4%).

• HealthBench Professional (Mythos 5): 66.0%. +1.3 over Claude Mythos Preview (64.7%), +9.1 over Claude Opus 4.8 (56.9%), +14.2 over GPT-5.5 (51.8%).

How and where to use it

Where it’s available

• The Claude app on Pro, Max, Team, and seat-based Enterprise plans at no extra cost through June 22 (phased rollout), then via purchased usage credits.

• The Claude API and consumption-based Enterprise from day one.

• Cloud-side: Claude Platform on AWS, Amazon Bedrock, Google Vertex AI, and Microsoft Foundry.

• Claude Mythos 5 is restricted to Project Glasswing participants, and a future trusted-access program will offer biology researchers a version with the bio/chem safeguards removed.

What it’s good at

• Long-horizon agentic coding and large-scale migrations

• Senior-level financial and analytical reasoning

• Scientific research workloads, vision tasks, and anyting where you’d previously reached for Opus 4.8 at maximum effort and still wanted more