Anthropic Released Claude Mythos (Fable 5) but Disabled Hacking and Biology Features

Anthropic has launched Claude Fable 5 and Claude Mythos 5, the company's most advanced AI models to date — but it deliberately disabled hacking- and biology-related capabilities in the public Fable 5 build. The decision followed internal evaluations showing the unrestricted Mythos 5 version could rapidly find and exploit long-standing security flaws and accelerate biological research, prompting Anthropic to impose safety limits on the public release of its most powerful AI model.

Fable 5 is priced at $10 per million input tokens and $50 per million output tokens — half the rate of the prior Mythos Preview. Max plan subscribers get doubled Opus usage limits through June 22, after which they will need to top up with credits to continue. The capped public build can be tried directly on the Claude website.

Anthropic said its internal tests showed the unrestricted system could discover and exploit vulnerabilities across major operating systems and software, including Linux, macOS, Firefox and OpenBSD. In one instance it exposed a 27‑year‑old OpenBSD flaw and a 16‑year‑old FFmpeg vulnerability. In lab exercises simulating biological research, non‑specialists using Mythos completed tasks in about 16 hours that the company estimates would otherwise take experts 40 to 95 days, and those sessions produced thousands of previously unknown security issues.

Benchmarks from Anthropic indicate Fable 5 outperforms Opus 4.8 and GPT‑5.5 on programming and agent‑style tasks, sustaining autonomous activity longer, preserving context across longer conversations, and handling images, graphs, tables and screenshots more effectively. Anthropic also notes Fable 5 exceeds the capabilities of its prior public releases.

The company is keeping the full‑strength Mythos 5 limited to a small set of cybersecurity partners while it works to mitigate risks and patch vulnerabilities before any wider rollout.

Do you think limiting access to advanced AI models is justified, or should open availability take priority over potential risks?