Anthropic officially released two AI models, Claude Fable 5 and Claude Mythos 5, on June 9, 2026. This is the first time that the company has brought "Mythos level"(mythical level) capabilities to the public eye, and it is also a very iconic one in the current big model competition."dual-track" release.
One model, two safety postures
Fable 5 and Mythos 5 share the same underlying architecture and are essentially two "clones" of the same model. Fable 5 is open to ordinary users and enterprises and is the "Claude model with the strongest publicly available capabilities" officially claimed by Anthropic; while Mythos 5 is only open to audited network security defenders, critical infrastructure providers and specific scientific research institutions through the Project Glasswing project, removing security restrictions in some highly sensitive areas.
thisThe core of the "one model, two versions" strategy lies in the differences in safety guardrails. Fable 5 has a built-in new classifier system that conservatively intercepts high-risk requests involving cybersecurity, biochemistry, and model distillation. Once triggered, it is automatically downgraded to Claude Opus 4.8 for processing. Anthropic says such "fallbacks" occur in less than 5% of sessions. Mythos 5 has lifted restrictions in these areas to unleash its full offensive, defensive and scientific research potential.
Fable 5: Publicly available "Strongest Claude"
In software engineering,The performance of Fable 5 is particularly outstanding. In SWE-Bench Pro (Real Software Engineering Task Benchmark), Fable 5 reached 80.3%, significantly ahead of Claude Opus 4.8 (69.2%), GPT-5.5 (58.6%) and Gemini 3.1 Pro (54.2%). In Cognition's FrontierCode difficult programming benchmark, Fable 5 far outperformed Opus 4.8 (13.4%) and GPT-5.5 (5.7%) with a score of 29.3%.
payment giant Stripe's early test case was convincing: Fable 5 completed a global migration of a 50-million-line Ruby code base in one day, a task that traditionally took the entire engineering team more than two months. Anthropic emphasized that Fable 5 's lead is becoming more evident in "longer and more complex tasks" and its autonomous running time exceeds that of any previous Claude model.
In terms of knowledge work and visual understanding,Fable 5 excels in financial reasoning, document analysis, and understanding of charts and tables. It can even reconstruct the source code of web applications based on screenshots alone. Its context window supports "millions of tokens", and in persistent memory tests (such as the game "Slaughter Spires "), Fable 5 reaches the endgame three times as often as Opus 4.8.
In terms of pricing,Fable 5 's API price is $10 per million for input and $50 per million for output, which is twice the price of Opus 4.8, but less than half the price of previous Mythos Preview. Until June 22, Pro, Max and Team subscribers will be available for free.
Mythos 5: The "Cybersecurity Myth" of Restricted and Open
The release of Mythos 5 continues Anthropic's "prudent openness" governance approach to high-capacity models. As early as April 2026, Anthropic refused to release it to the public because Mythos Preview demonstrated the ability to "independently discover and exploit zero-day vulnerabilities." It only cooperated with about 40 organizations through Project Glasswing. This time, Mythos 5 partners have expanded to approximately 200 institutions, covering 15 countries.
In Cybersecurity Benchmarks In ExploitBench, Mythos 5 scored 78.0%, which was significantly higher than Mythos Preview (69.0%), Opus 4.8 (40.0%) and GPT-5.5 (34.0%). Anthropic calls this "the strongest cybersecurity modeling capability in the world."
In the field of life sciences,Mythos 5 also demonstrates amazing scientific research acceleration capabilities. In-house protein design experts have used this model to speed up some processes by about 10 times, and 9 of the 14 protein goals have produced strong candidates. In blind testing of molecular biology hypotheses, scientists prefer the results given by Mythos 5 about 80% of the time.
Security disputes and industry significance
The release of Fable 5 was not without controversy. Although Anthropic spent more than 1000 hours of vulnerability bounty testing and failed to find a "universal jailbreak" path, the British AI Security Institute (AISI) has found some breakthroughs in initial testing. In addition, although conservative security classifiers control the accidental injury rate below 5%, they may cause experience losses to users involved in legitimate scientific research needs such as biology and chemistry.
From an industry perspective,Anthropic's "dual-track" release has exemplary significance: on the one hand, it commercializes cutting-edge capabilities through Fable 5, competing head-on with OpenAI's GPT-5.5 and Google's Gemini 3.1 Pro; on the other hand, it explores "controlled openness" through Mythos 5 and Project Glasswing. The governance model attempts to find a balance between capacity release and risk management and control. With Anthropic and OpenAI both submitting confidential IPO documents, this game between the "strongest model" and the "safest release" is becoming one of the most noteworthy story lines in the AI industry in 2026.