Anthropic releases Mythos-class AI to public with Claude Fable 5 launch

Anthropic has released its top-tier Mythos-class AI model to the general public, ending a delay driven by security concerns.

The company on Wednesday announced the launch of Claude Fable 5 and a security-focused variant, Claude Mythos 5 — the most capable models Anthropic has released to date.

The two models share the same underlying architecture, but Fable 5 adds safety guardrails for sensitive domains and is intended for general users. Mythos 5 will be distributed selectively to institutions participating in Glasswing, Anthropic's security consortium.

When Fable 5 receives a cybersecurity query that could be misused, the request is routed to Opus 4.8 — the next model down — which handles the response instead, and the user is notified.

"We have long hoped to develop new safeguards strong enough to prevent misuse so that we could make Mythos-class capabilities available to all users," Anthropic said. "We have been improving those safeguards over the past several months, and they are now sufficiently robust."

The company added that some benign requests may still be blocked, but that early data showed more than 95 percent of sessions were not affected.

Mythos 5 will be made available only to verified organizations through Project Glasswing, Anthropic's security partnership program. South Korean participants include SK Telecom, SK Hynix, Samsung Electronics and the Korea Internet & Security Agency (KISA).

Anthropic also introduced a new policy of retaining data generated by Fable and Mythos models for up to 30 days, with the aim of detecting misuse attempts and refining its safety guardrails.

Both Claude Fable 5 and Mythos 5 are priced at $10 per million input tokens and $50 per million output tokens. The models are available starting Wednesday and will be offered to existing paid subscribers at no extra charge through June 22, after which a separate fee will apply.

Anthropic said it plans to fold Fable 5 back into existing subscription tiers once it has secured sufficient server capacity.

The company said the new models deliver state-of-the-art performance across nearly all benchmarks.

In the ExploitBench evaluation, Mythos 5 scored 78 percent, surpassing the Mythos preview released two months ago, which scored 69 percent.

On the Humanity's Last Exam (HLE) benchmark, which measures doctoral-level intelligence across disciplines, Mythos 5 scored 59 percent without tool use — topping the Mythos preview's 56.8 percent, which had been the first model to clear the 50 percent threshold. Anthropic said Fable 5's safety guardrails prevented a comparable performance measurement for that model.