Anthropic has launched Claude Fable 5 for general use and Claude Mythos 5 for a restricted group of users. The two systems share the same underlying model.
Fable 5 is being released with safeguards that route some prompts to Anthropic's next-most-capable model, Claude Opus 4.8, rather than answering directly. Mythos 5 is available to a small group of cyber defenders and infrastructure providers, with some of those limits removed through Project Glasswing, which Anthropic runs with the US government.
The launch reflects Anthropic's effort to widen access to a more powerful model while limiting use in areas it considers especially sensitive, including cybersecurity, biology, chemistry and attempts to extract model behaviour for training rival systems.
Anthropic said the safeguards were tuned conservatively, meaning some harmless requests would be caught, but fallback to Opus 4.8 occurred in fewer than 5% of sessions on average. In the remaining sessions, users interact with Fable 5 without interruption.
Anthropic is charging USD $10 per million input tokens and USD $50 per million output tokens for both models. That is less than half the price of Claude Mythos Preview, an earlier restricted model made available through the same programme.
Safety controls
The main distinction between Fable 5 and Mythos 5 is the safeguards, not the core model. Anthropic said separate AI classifiers screen prompts for potential misuse, including jailbreak attempts, and trigger a handover to Opus 4.8 when a request falls into protected categories.
Those categories include offensive cybersecurity work, biology and chemistry topics that could create dual-use risks, and large-scale distillation efforts designed to copy the model's behaviour. Users are told when that fallback occurs.
Anthropic said it subjected the controls to internal and external testing. An external bug bounty produced no universal jailbreaks in more than 1,000 hours of testing, though the company added that no system was likely to prevent all such attempts indefinitely.
Anthropic also said one external partner found that Fable 5 complied with no harmful single-turn requests related to planning a cyberattack, exploit development or defence evasion, including prompts using public jailbreak methods.
Performance claims
Anthropic described Fable 5 as its strongest generally available model to date and said it performed well across software engineering, analytical work, vision tasks, memory and scientific research. It said the gap widened on longer and more complex work.
In software engineering, Anthropic cited early testing by Stripe. "Fable 5 compressed months of engineering into days," it said Stripe reported, adding that the model completed a codebase-wide migration across a 50-million-line Ruby codebase in a day, a task it said would otherwise have taken a team more than two months to do manually.
For analytical work, Anthropic said the model achieved the highest score on Hebbia's finance benchmark for senior-level reasoning and performed strongly in document-based reasoning, chart and table interpretation, and problem solving. It also said IMC found the model performed well in trading-analysis tests spanning factual lookup, conceptual reasoning, root-cause analysis and expected-value analysis.
On vision tasks, Anthropic said Fable 5 could extract precise figures from scientific charts and rebuild web application source code from screenshots. It also highlighted game-playing tests, saying the model completed Pokémon FireRed using only raw screenshots and no navigation aids, whereas earlier versions needed more support.
Science and cyber use
Anthropic reserved Mythos 5 for users in higher-trust settings because of what it described as stronger cyber and scientific performance. The company said the model had "the strongest cybersecurity capabilities of any model in the world" and was already being used by cyber defenders working on critical software.
It also pointed to internal work in drug design and molecular biology. Anthropic said its protein design researchers found Mythos 5 could speed up parts of the drug design process by around ten times and, in some cases, match or exceed skilled human operators in selecting binding sites, running design tools and recovering from failures without human help.
In one genomics project, Anthropic said Mythos 5 worked largely autonomously for more than a week, assembling single-cell data across 138 animal species and training a custom machine learning model that outperformed a recent model published in Science despite being 100 times smaller.
Anthropic also said its scientists preferred Mythos-generated molecular biology hypotheses about 80% of the time in blinded head-to-head comparisons with Opus-class models, and that several had moved to experimental evaluation.
Data policy
Alongside the launch, Anthropic is changing how it handles business customer data for Mythos-class systems and future models at similar or higher levels. The company will require 30-day retention for all traffic on those models across both first-party and third-party services.
Anthropic said it would not use that data to train new Claude models or for non-safety purposes. It added that the retention period was intended to help detect complex attacks, including new jailbreaks and behaviour spread across multiple requests, and to reduce false positives in the filtering system.
Access to Mythos 5 remains limited to existing Glasswing partners and, later, a small number of biology researchers, who will receive versions with biology and chemistry safeguards removed but cyber safeguards still in place. Anthropic said, "We expect demand for Fable 5 to be very high, and difficult to predict."