Anthropic unveiled Claude Fable 5 and Claude Mythos 5 on Tuesday, launching its most powerful AI models to date while introducing new safeguards meant to limit misuse in cybersecurity, biology and model distillation.
The company said in a blog post that Claude Fable 5 is now available for general use, describing it as a “Mythos-class” model that outperforms its previous widely released systems across most tested benchmarks, including software engineering, knowledge work, vision, scientific research and long-running autonomous tasks.
Claude Mythos 5, which uses the same underlying model, will initially be restricted to a small group of cyber defenders and infrastructure providers through Project Glasswing, an initiative Anthropic is running in collaboration with the U.S. government to give trusted organizations access to stronger AI tools for defensive work.
Anthropic opens access, but keeps guardrails tight
Anthropic said the launch reflects the challenge of making frontier AI more widely available while reducing the risk that advanced capabilities could be used for harmful cyber or biological activity.
To support a broader release, Fable 5 includes new safety classifiers that detect sensitive requests and route them to Claude Opus 4.8, Anthropic’s next-most-capable model, rather than allowing Fable 5 to respond directly.
The company said the system covers areas such as offensive cybersecurity, biology and chemistry, and attempts to distill the model’s capabilities into competing AI systems.
Anthropic said the fallback system is intended to preserve most of Fable 5’s benefits for everyday users while reducing the chance that its stronger capabilities could assist cyberattacks, risky biological research or the spread of near-frontier models without comparable controls.
Stronger results across code, vision and research
Anthropic positioned Fable 5 as a major leap for complex, multi-step work, saying its advantage over earlier Claude models grows as tasks become longer and more demanding.
In software engineering, the company cited early testing from Stripe, which said Fable 5 completed a codebase-wide migration in a 50-million-line Ruby project in one day, compared with more than two months of work by a full team manually.
Anthropic also said Fable 5 performed strongly on Cognition’s FrontierCode evaluation while using fewer tokens than earlier Claude models.
The company said the model also improved in analytical work, including finance reasoning, document-heavy analysis, chart and table interpretation, and problem solving.
In vision, Anthropic said Fable 5 can extract precise numbers from scientific figures and rebuild web app source code from screenshots.
One demonstration showed Fable 5 completing Pokémon FireRed using only raw visual input, without maps, navigation tools or extra game-state information, a task earlier Claude models needed more support to handle.
Mythos 5 stays reserved for trusted users
While Fable 5 is available broadly, Mythos 5 is being limited to users who need fewer restrictions in specific domains. Anthropic said the model has the strongest cybersecurity capabilities of any AI system and will upgrade access for existing Project Glasswing participants.
The company plans to expand Mythos 5 through a trusted access program for cybersecurity organizations, alongside a separate biology program for selected researchers.
In the biology version, users would receive access with biology and chemistry safeguards removed, while cyber restrictions would remain in place.
Highlighting Mythos 5’s stronger reach in biology, Anthropic said the model is already showing potential beyond benchmark tests, particularly in drug discovery, molecular research and large-scale genomics.
Anthropic said internal protein design experts used Mythos 5 to speed up parts of the drug design process by about tenfold.
The company added that Mythos 5 helped identify strong drug-design candidates for nine of 14 protein targets, generated molecular biology hypotheses preferred by Anthropic scientists, and conducted largely autonomous genomics research across millions of cells from 138 animal species.
New retention rules and pricing
Anthropic also announced a new data retention policy for Fable 5, Mythos 5 and future models with similar or higher capabilities. The company said it will require 30-day retention for all Mythos-class model traffic across first- and third-party surfaces.
Anthropic said the retained data will not be used to train new Claude models or for non-safety purposes. The policy is designed to help detect complex attacks, identify jailbreak attempts across multiple requests and reduce false positives in its safeguards.
Both Claude Fable 5 and Claude Mythos 5 are priced at $10 per million input tokens and $50 per million output tokens.
Fable 5 is available through the Claude API and consumption-based Enterprise plans from launch. Anthropic said users on Pro, Max, Team and seat-based Enterprise plans will have access at no extra cost through June 22, with usage credits required from June 23 unless the company extends the initial access window.
