Anthropic introduces two faces of the same innovation today: Claude Fable 5, a Mythos-class model tuned for general use with strong safeguards, and Claude Mythos 5, which runs on the same engine but relaxes some restrictions for trusted cyber defenders and researchers. What does this mean for developers, researchers, and curious users like you? Let's break it down.
What are Fable 5 and Mythos 5
Claude Fable 5 is a Mythos-class model adapted for general use: it delivers the best performance Anthropic has publicly offered so far, standing out in programming, knowledge work, computer vision, and research tasks. In long, complex tests, its advantage over earlier models grows.
Claude Mythos 5 uses the same base model but lifts safeguards in specific areas (mainly cybersecurity and some scientific uses). Mythos 5 is being rolled out first through Project Glasswing with government and security partners; the plan is to extend access via a trusted access program.
Capabilities that stand out
-
Software engineering: companies like Stripe reported tasks that used to take months compressed into days. In one migration of 50 million lines of Ruby code, the model did in a day what a team would have needed two months to finish.
-
Knowledge work and finance: in senior-reasoning and financial-analysis benchmarks Fable 5 leads, improving document interpretation, table handling, and expected-value analysis.
-
Vision: it can extract precise numbers from scientific figures and reconstruct app code from screenshots. It even outperformed prior models on visual control tasks in games with minimal guidance.
-
Memory and long context: it keeps focus over millions of tokens and improves with persistent notes; in deck-building style game tests its performance rose significantly compared to previous models.
-
Biology and protein design (Mythos 5): in internal tests it sped up stages of drug design by nearly 10×, running complex workflows without human assistance and producing promising candidates across several targets.
-
Genomics: it conducted automated research that produced a cell-identification model which outperformed a recent Science paper, using a much smaller model.
Safeguards: how they protect and what they limit
Anthropic introduces classifiers that detect attempts at harmful use. When queries are identified as related to cybersecurity, biology/chemistry, or as attempts to “distill” capabilities to train rival models, Fable 5's response is routed to Claude Opus 4.8.
In practice: more than 95% of sessions with Fable 5 do not require this routing. When it happens, the user is informed and receives an Opus 4.8 response instead of a flat denial.
Key points about the classifiers:
-
Cybersecurity: Mythos-class models can make it easier to find and exploit vulnerabilities; classifiers therefore block offensive and agentic tasks. Anthropic ran red-teaming and a bug-bounty program and did not find a universal "jailbreak" in those tests.
-
Biology and chemistry: because they can perform scientific tasks with dual-use risk, many queries in these areas currently fall back to Opus 4.8. Anthropic plans to broaden access for biomedical researchers via a trusted-access program.
-
Distillation: large-scale attempts to extract the model's knowledge to recreate it elsewhere will be detected and diverted.
Privacy and data retention
For Mythos-class models (Fable 5, Mythos 5 and similar future models) Anthropic requires 30-day retention of traffic on first- and third-party surfaces. This data will not be used to train new models or for purposes other than security. There are additional controls: human access to the logs is recorded and data is deleted after 30 days, except when retained briefly to investigate attacks.
Access, price and rollout
-
Price: 10 USD per million input tokens and 50 USD per million output tokens.
-
Availability: Claude Fable 5 is available globally today and can be used via the API with the identifier
claude-fable-5. -
Mythos 5 is initially limited to Glasswing partners; expansion will be gradual through trusted-access programs for cybersecurity and for biology.
-
Subscriptions: Fable 5 will be included at no extra cost in Pro, Max, Team, and Enterprise plans per seat from today until June 22. From June 23, use will require consumption credits unless the free window is extended due to capacity.
Risks and security outcomes
Anthropic acknowledges Mythos-class models are dual-use and has taken a conservative approach with classifiers, which can produce false positives and cause frustration. Still, the team reports:
- Automated tests show low misaligned behavior similar to Opus 4.8.
- The classifiers block progress on offensive tasks and have been subjected to external red-teaming.
- There's evidence of positive utility: from hardening critical software in Project Glasswing to scientific hypotheses that independent labs corroborated.
Does this mean the risk disappears? No. It means they're trying to balance access and protection, and they'll ask for more public collaboration to strengthen defenses.
What you can do now
-
If you're a developer, try it on long engineering and automation tasks; Fable 5 shines in projects that need to maintain context and sustained reasoning.
-
If you work in cybersecurity or critical infrastructure, track the Glasswing program and the trusted-access route to request Mythos 5.
-
If you're a biomedical researcher interested in advanced capabilities, prepare an application for the trusted-access biology program—spots will be limited.
-
If you use the public version and encounter false positives or unexpected behavior, report them: that feedback helps refine the classifiers.
The arrival of Fable 5 and Mythos 5 is a practical reminder: the capability frontier keeps moving, and with it the responsibility. It's not just a new powerful tool; it's a social and technical experiment about how to deploy high-impact AI without putting the world at risk.