Claude Fable 5 Is Live - Mythos-Class AI Goes Public at 80.3% SWE-Bench Pro with Hard Safety Limits and a June 22 Free Window

What Fable 5 Is and How It Differs from Mythos Preview

Fable 5 is Anthropic's answer to the tension that has defined Claude Mythos Preview since its April 2026 debut: the most capable model they have ever built is also the one most dangerous to release widely. Mythos Preview was restricted to ~200 organizations through Project Glasswing specifically because it could autonomously discover and chain zero-day exploits across major operating systems and browsers. Releasing that capability without guardrails would be, in Anthropic's framing, irresponsible.

Fable 5 solves this by adding a safety classifier layer. When a query touches cybersecurity, biology, chemistry, or distillation at the level where Mythos's dual-use capabilities become a concern, Fable 5 routes the query to Claude Opus 4.8 instead. The user gets an answer; it just comes from the model whose capabilities in that domain are less dangerous. Anthropic says this safeguard triggers in under 5% of sessions on average. For the vast majority of use cases - coding, research, writing, vision, data analysis, agentic workflows - Fable 5 responds at full Mythos-class capability.

Anthropic's primary announcement states directly: "Today we're launching Claude Fable 5: a Mythos-class model that we've made safe for general use. Fable 5's capabilities exceed those of any model we've ever made generally available." This is Anthropic's first admission in a public launch that their most capable model class is now available to everyone - not just vetted security partners.

Benchmarks - How Fable 5 Compares

The SWE-Bench Pro figure is the one that matters most for developers. 80.3% is a significant step above Opus 4.8 (69.2%) and a large lead over GPT-5.5 (58.6%) on hard, real-world multi-file engineering problems. Anthropic's primary announcement frames the competitive position clearly: "The longer and more complex the task, the larger Fable 5's lead over our other models." This is not a generalist improvement - it is specifically designed to excel where other models struggle most.

Pricing and Model Details

At $10/$50 per million tokens, Fable 5 is exactly 2x the price of Opus 4.8 ($5/$25). MacRumors confirmed this is "less than half the price of the Claude Mythos Preview" - the restricted Glasswing access tier was considerably more expensive. The 128K output token limit is a practical advantage over Opus 4.8 for long autonomous tasks - generating large amounts of code, documentation, or research output in a single session without hitting the ceiling.

The June 22 subscription cut-off is the detail most Pro and Max subscribers need to act on. From today until June 22, Fable 5 is included in all paid plans at no additional cost. On June 23, Anthropic removes it from subscription plans and usage requires credits. This is a deliberate capacity management strategy - they cannot yet serve Fable 5 at the volume of their full subscription base without degrading quality, so they are staging the rollout while infrastructure scales. Anthropic committed to restoring Fable 5 as a standard subscription feature once capacity is sufficient.

The Safety Architecture - What Gets Blocked and Why

Fable 5's safety architecture is the most significant part of the announcement from a policy perspective. Anthropic is releasing a model they describe as capable of autonomously discovering and chaining zero-day exploits - and releasing it publicly. They are doing so with a classifier that routes sensitive queries to Opus 4.8 instead. The categories that trigger routing:

The classifier returns stop_reason: "refusal" as an HTTP 200 response when routing occurs - not a hard error. Developers using the Fable 5 API should handle this gracefully in their applications. Importantly: if the refusal returns with no output generated, you are not billed for the request. This is a new billing behavior confirmed in Anthropic's platform documentation. GitHub Copilot deployments of Fable 5 require up to 30 days of prompt-and-output retention to operate the safety classifiers - this is off by default in Copilot admin settings and must be explicitly enabled by the organization admin.

Mythos 5 - The Still-Restricted Tier

Alongside Fable 5, Anthropic also launched Mythos 5 - an updated version of the fully capable Mythos model without the Fable 5 safety routing. Mythos 5 is NOT publicly available. It is available only to organizations that already have access to Mythos Preview through Project Glasswing - currently approximately 200 organizations across 15+ countries including Apple, AWS, Google, Cisco, Microsoft, IBM, Cloudflare, Mozilla, Palo Alto Networks, NATO, and EU ENISA. TechCrunch confirmed that Apple is one of the Glasswing partners receiving Mythos 5.

For everyone else, Fable 5 is the publicly available Mythos-class model - the same underlying capability but with the safety routing layer that prevents the most dangerous dual-use applications. Anthropic describes Mythos 5 and Fable 5 as sitting "above our Opus class in capability" - a new tier in the model hierarchy above Opus, Sonnet, and Haiku. For the full Glasswing context, see our Project Glasswing expansion article.

What This Means for Developers

For developers building on Claude Code, the 80.3% SWE-Bench Pro figure is the most relevant benchmark. That is an 11-point improvement over Opus 4.8 on hard real-world engineering problems. For teams running long autonomous coding sessions - the kind where Claude Code runs for hours on a codebase-scale refactor - Fable 5's extended autonomous work capability is the key practical improvement. The 128K output token limit means fewer interruptions on large generation tasks. For the full comparison of coding agents following this launch, see our updated Codex vs Claude Code comparison.

For API users: add claude-fable-5 to your model rotation today. The model is available on the Claude API, AWS Bedrock, and GitHub Copilot. Handle stop_reason: "refusal" gracefully in your application - it is a normal response for queries that trigger the safety routing, not an error. You are not billed for these responses if no output was generated. For subscription users on Pro, Max, Team, or Enterprise: you have until June 22 to use Fable 5 at no extra cost. Test it on your hardest use cases now. For all model news and pricing changes, see the June 2026 AI news calendar and May 2026 archive.

Frequently Asked Questions