Claude Mythos gets a public face: Meet Fable 5 with built-in guardrails for the rest of us

Anthropic has released two new AI models: Claude Fable 5, available to the general public, and Claude Mythos 5, reserved for a small group of vetted partners. Both share the same underlying model, but Fable 5 comes with safety restrictions that block responses in high-risk areas. The launch marks the first time Anthropic has made its most capable model class broadly accessible.

Same model, different access

Fable 5 and Mythos 5 are built on the same foundation. The difference lies in what they are allowed to do. Fable 5 includes classifiers, which are separate AI systems that monitor incoming requests and intercept anything that touches cybersecurity, biology, chemistry, or attempts to copy the model’s capabilities. Those requests are automatically rerouted to Claude Opus 4.8, an older and less powerful model. Anthropic says this fallback affects fewer than 5% of sessions.

Mythos 5 removes those restrictions in specific areas. It is being deployed through Project Glasswing, a collaboration with the US government that launched in April, and is currently limited to cybersecurity partners and, beginning today, a small group of biology researchers. Anthropic describes it as the most capable cybersecurity model in the world.

Anthropic’s head of product management, Diane Penn, told WIRED that the fallback approach emerged as the most practical option after months of testing and user feedback. “We’re trying to make improvements in a way that’s beneficial, even if we don’t have the perfect solution for every use case to start,” she said.

What the models can do

Anthropic claims Fable 5 outperforms every model the company has previously released to the public, and says it leads across nearly all capability benchmarks it was tested on. Performance gains are most pronounced on long and complex tasks.

In software engineering, payment processor Stripe reported that the model compressed months of work into days. It completed a migration across a 50-million-line codebase in a single day, a task that Stripe estimated would have taken a full team more than two months by hand.

On vision tasks, Fable 5 can reconstruct a web application’s source code from screenshots alone. It also completed the video game Pokémon FireRed using only visual input, without the additional tools that earlier Claude models required.

Mythos 5 shows particularly strong results in scientific work. Anthropic’s internal protein design team said the model accelerated parts of the drug design process by around ten times. In one test, it matched or outperformed skilled human operators while working without human assistance. Nine out of 14 protein targets produced viable drug design candidates now under investigation.

In genomics, Mythos 5 worked largely autonomously for over a week. It compiled single-cell data from 138 animal species, then designed and trained its own machine learning model. Anthropic says the result outperformed a recently published model from the journal Science, despite being 100 times smaller.

Anthropic also claims Mythos 5 is the first model to consistently generate novel and credible scientific hypotheses. In internal blind comparisons, researchers preferred its molecular biology hypotheses over those from Opus-class models roughly 80% of the time. One hypothesis about an E. coli protein mechanism was later corroborated by an independent research group.

Safety measures and data retention

Anthropic red-teamed the classifiers extensively before launch. An external bug bounty produced no universal jailbreaks across more than 1,000 hours of testing. External red-teaming organisations also failed to find universal jailbreaks, though the UK AI Safety Institute made partial progress in an initial testing window.

Alongside the launch, Anthropic introduced a new data retention policy. All traffic on Mythos-class models will be stored for 30 days, even for business customers who previously held zero-retention agreements. The company says this data will not be used for training and will only be used to detect novel attacks and reduce false positives from the classifiers.

Anthropic acknowledges the classifiers are currently tuned conservatively and will sometimes block harmless requests. The company says it plans to reduce these false positives over time.

Pricing and availability

Both models are priced at $10 per million input tokens and $50 per million output tokens. Anthropic says this is less than half the price of the earlier Mythos Preview, but it is roughly double the cost of Claude Opus 4.8.

Fable 5 is available now through the Claude API and usage-based Enterprise plans. The rollout for subscription plans follows a staged approach:

• Until June 22: Fable 5 is included at no extra cost on Pro, Max, Team, and seat-based Enterprise plans.
• From June 23: Access requires usage credits on those plans.
• Later: Anthropic plans to restore Fable 5 as a standard part of subscription plans once capacity allows.

Mythos 5 remains restricted to Project Glasswing partners and select biology researchers for now. Anthropic says it plans to expand access through a broader trusted access program in coordination with the US government.

Sources

• Claude Fable 5 and Claude Mythos 5 – Anthropic
• Anthropic releases Claude Fable 5 and Mythos 5 with major gains in coding and science – The Decoder
• Anthropic releases Claude Fable, a version of Mythos, days after warning AI is becoming too dangerous – TechCrunch
• Anthropic’s new Claude Fable 5 is the same base model as Mythos but with guardrails attached – ZDNET
• Anthropic Offers Mythos Upgrade for Cyber Partners and a ‘Safe’ Version for the Rest of You – WIRED