Anthropic launches Claude Fable 5 with new safeguards

Anthropic has launched Claude Fable 5 as a generally available Mythos-class model with new safety controls. According to the company, Fable 5 can handle longer and more complex tasks than prior Claude models.

The release also includes Claude Mythos 5 for selected cyberdefenders and infrastructure providers.

Claude Fable 5 enters general release

Claude Fable 5 is now available to users through Claude products and the Claude API. Developers can access the model through the claude-fable-5 API identifier. Anthropic said Fable 5 performs strongly in software engineering, knowledge work, vision, and scientific research. 

Additionally, Anthropic has noted that the model has its largest lead on longer and more complex tasks. The company said Fable 5 can work autonomously for longer periods than earlier Claude models. It also said the model can stay focused across millions of tokens in long-running tasks.

In software testing, Stripe reported that Fable 5 completed a large Ruby migration in one day. Stripe said the same migration would have taken a team more than two months by hand. Anthropic also cited strong results on finance, vision, memory, and scientific research tasks. However, the company said the release required extra controls because of the model’s capabilities.

Safeguards route some queries to Opus 4.8

Anthropic said some Fable 5 requests will fall back to Claude Opus 4.8 instead. The fallback applies to selected cybersecurity, biology, chemistry, and distillation-related queries. The company said the safeguards trigger in less than 5% of sessions on average. It also said more than 95% of Fable sessions involve no fallback.

Anthropic said the safeguards may catch harmless requests because it tuned them conservatively. The company said it wants to reduce false positives after launch. The company added new classifiers to detect potential misuse and jailbreak attempts. 

These systems prevent Fable 5 from responding directly to flagged requests. Anthropic said the cybersecurity controls cover exploitation and other offensive cyber tasks. The company said biology and chemistry safeguards cover many requests because of dual-use risks.

Mythos 5 access starts with trusted groups

Anthropic also launched Claude Mythos 5 for a smaller group of approved users. The company said Mythos 5 uses the same underlying model as Fable 5. Mythos 5 starts through Project Glasswing in cooperation with the U.S. government. Anthropic said the program includes cyberdefenders and critical software infrastructure providers.

The company said Mythos 5 lifts some safeguards for approved cybersecurity users. It also plans a trusted access program for selected biology researchers and companies. Anthropic said business customer traffic on Mythos-class models will face 30-day retention. The company said it will not use this data to train new Claude models.

The company set pricing for Fable 5 and Mythos 5 at $10 per million input tokens. It also set output pricing at $50 per million tokens. Fable 5 is included on Pro, Max, Team, and seat-based enterprise plans through June 22. Starting June 23, Anthropic said use will require credits unless capacity allows an extension.