Anthropic Opens Mythos-Class AI to the Public With Claude Fable 5 Safeguards

Anthropic has opened public access to Claude Fable 5, a frontier model based on the same core system behind the more restricted Claude Mythos 5. The release is notable because Mythos-class capability has attracted close attention from governments, banks, cybersecurity teams, and other institutions worried about how stronger AI systems could be misused in high-risk domains.

Fable 5 is Anthropic's answer to that tension: release more capability to general users, but wrap it in conservative safety routing. The company is positioning the model as its strongest public Claude system so far, especially for software engineering, research, advanced reasoning, vision work, and long tasks that require a model to maintain context over extended sessions.

Mythos 5 remains limited to a smaller trusted-access group. It shares the same base model family, but has fewer restrictions for approved cyberdefense organizations, infrastructure providers, and specialist research partners.

One Base Model, Two Safety Profiles

The central difference between Fable 5 and Mythos 5 is not raw intelligence. It is access control.

Fable 5 includes additional classifiers that watch for high-risk requests in areas such as offensive cybersecurity, biology, chemistry, and model distillation. When those classifiers trigger, the request is routed away from Fable 5 and handled by Claude Opus 4.8 instead. Users are told when this fallback happens. Anthropic says the safeguards trigger in fewer than 5% of sessions on average, but they have been tuned conservatively, meaning some harmless requests may be caught too.

That tradeoff is important. Rather than holding the model back entirely, Anthropic is trying a middle path: keep the stronger model available for most work, but move riskier requests into a stricter model lane.

Mythos 5 removes some of those restrictions for approved users. Its initial deployment is tied to Project Glasswing, Anthropic's trusted-access effort for cyberdefenders and critical software infrastructure teams. The company plans to widen access later through a broader trusted program.

Stronger Long-Horizon Coding and Agentic Work

Anthropic is emphasizing Fable 5's ability to operate over longer tasks with less hand-holding. Early testing highlights include codebase-scale migrations, production-grade coding evaluations, and multi-step agentic development work that previously required more scaffolding or more human intervention.

One of the most striking examples involves large-scale software maintenance: Fable 5 reportedly handled a migration across a 50-million-line Ruby codebase in a single day, a job estimated to require more than two months from a full engineering team if done manually.

The model also appears designed for the increasingly competitive agentic coding market, where models are judged less by isolated benchmark answers and more by whether they can plan, execute, recover from mistakes, and finish real work. That puts Fable 5 in direct competition with other high-end coding and agent models, including the frontier systems covered in our GPT-5.5 release analysis.

Vision, Memory, and Scientific Reasoning Improve

Beyond code, Fable 5 shows major upgrades in visual understanding and long-context behavior. Anthropic says the model can extract precise values from dense scientific figures, reason over complex charts and tables, and even reconstruct web application source code from screenshots.

The company also points to improved persistence across extremely long tasks. Fable 5 can make better use of notes and file-based memory, allowing it to maintain strategy and self-correct over extended runs. In game-based internal evaluations, persistent memory helped Fable 5 substantially more than it helped Opus 4.8.

Mythos 5 is being highlighted most strongly for scientific and security work. Anthropic says internal protein-design specialists used Mythos 5 to accelerate parts of drug-design workflows by roughly an order of magnitude. In one test, the model selected binding sites, chose tools, ran protein-design workflows, and recovered from failures with little direct human help.

The model also generated molecular biology hypotheses that Anthropic scientists preferred over previous Opus-class outputs in blinded comparisons, and it reportedly conducted a multi-day genomics research workflow involving millions of cells across 138 animal species.

Pricing Lands Below Mythos Preview

Both Fable 5 and Mythos 5 are priced at $10 per million input tokens and $50 per million output tokens. That is less than half the price of the earlier Mythos Preview tier, making the new models substantially cheaper for heavy frontier-model workloads.

The pricing also suggests Anthropic wants Fable 5 to be used at scale, not just treated as a rare premium model for occasional high-stakes prompts. If the fallback system works as intended, Fable 5 could become the default choice for difficult coding, analysis, and multimodal tasks while Mythos 5 remains gated for the small number of users who need sensitive domain capability.

30-Day Retention Adds Another Safety Layer

Anthropic is also requiring 30-day retention for traffic on Mythos-class models across its own products and third-party surfaces. The retained data is meant for safety work rather than model training, including detecting complex jailbreak attempts, spotting attacks that span many requests, and reducing false positives in the safeguard system.

That retention requirement is likely to matter for enterprise buyers. Banks, government agencies, and regulated companies want stronger models, but they also need a clear account of what gets logged, who can access it, and when it is deleted. Anthropic says human access to retained data is logged and that the data is deleted after 30 days in almost all cases.

Why This Launch Matters

The Fable/Mythos split is a useful signal for where frontier AI deployment is heading. Model providers increasingly want to ship stronger systems quickly, but the most capable models are also more useful for dual-use work in cyber, biology, and automated research.

Anthropic's answer is not a single public model with uniform rules. It is a tiered release strategy: general users get most of the capability most of the time, while specific risky domains are routed through extra controls or reserved for trusted-access programs.

That model may become more common as frontier systems improve. The practical question is whether users accept occasional false positives in exchange for broader access to more capable systems — and whether trusted-access programs can scale without becoming opaque gatekeeping layers around the most powerful AI tools.