Anthropic releases first public Mithos-class AI with safety guardrails in Fable 5

Anthropic, the AI startup pursuing an initial public offering, has for the first time publicly released an AI model at its top-tier "Mithos" capability level — a class it had previously withheld from general release. The consumer-facing version, however, carries separate safety guardrails over concerns about potential misuse in cybersecurity attacks and biological or chemical weapons development.

Anthropic announced Tuesday that it was launching two models: Claude Fable 5, intended for general users, and Claude Mithos 5, a security-focused variant.

The company said the name Fable derives from the Latin word "Fabula," meaning myth, and that Fable 5 is in effect the same model as Mithos 5.

The key difference lies in the safety restrictions applied to Fable 5. When a user submits a query that could relate to cyberattacks, bioweapon development or chemical misuse, Fable 5 does not process the request directly. Instead, it routes the response through Opus 4.8, the next model down, and notifies the user accordingly.

The same restrictions apply to requests suspected of attempting to extract capabilities from competing AI models through so-called distillation.

"We designed the guardrails conservatively to enable a safe and rapid launch," Anthropic said, adding that while some harmless requests may be blocked, the restrictions activate in fewer than 5 percent of all sessions.

Mithos 5, which carries no such restrictions, will be made available only to institutions verified through a security consortium called Project Glasswing.

Samsung Electronics, SK Hynix, SK Telecom and the Korea Internet & Security Agency are understood to be among the participants in Project Glasswing, meaning those organizations may be able to secure access to Mithos 5.

Anthropic also introduced a new data policy under which data generated through use of Fable 5 and Mithos 5 will be retained for 30 days to help detect new attack techniques and analyze false positives.

Both models significantly outperformed existing publicly available models on benchmarks.

On ExploitBench, which measures cybersecurity capabilities, Mithos 5 scored 78 percent — surpassing GPT-5.5 at 34 percent, Opus 4.8 at 40 percent, and the Mithos Preview released two months ago at 69 percent.

On the Humanity's Last Exam, which tests doctoral-level problem-solving, Mithos 5 scored 59 percent, topping the Mithos Preview's score of 56.8 percent — itself the first AI model to break the 50 percent threshold.

Coding performance also improved. On Terminal Bench 2.1, which evaluates terminal-based programming ability, Mithos 5 scored 88 percent, edging out GPT-5.5 at 83.4 percent.

On SWE-Bench Pro, which assesses general software development capability, Fable 5 scored 80.3 percent, well ahead of GPT-5.5 at 58.6 percent and Google Gemini 3.1 Pro at 54.2 percent.

On GDPval-AA, a benchmark for knowledge work performance, Fable 5 scored 1,932 points, surpassing GPT-5.5 at 1,769 points and Gemini 3.1 Pro at 1,314 points.

Fable 5 is available starting Wednesday and will be offered to existing paid subscribers at no additional charge through June 22, after which a separate pricing plan will apply.

Anthropic said it is considering reintegrating Fable 5 into existing subscription products once sufficient server capacity is secured.