On June 9, 2026, Anthropic released Claude Fable 5, the first Mythos-class model available to the general public. At the same time, the company introduced Claude Mythos 5 for a narrow group of cybersecurity specialists. Under the hood, it is the same model; the only difference is in the protective mechanisms.
One model, different protections
According to Anthropic, Fable 5 outperforms any model the company has previously released and shows the best results on almost all tested benchmarks: software development, analytical work, computer vision, and scientific research. The longer and more complex the task, the more noticeable the gap.
Without restrictions, the model’s cybersecurity skills could be used for harm. Therefore, for a number of topics, Fable 5 automatically routes requests to the less powerful Claude Opus 4.8 model and notifies the user about it. The protections are configured conservatively and sometimes trigger on harmless requests, but on average this happens in less than 5% of sessions.
Mythos 5 is intended for specialists defending against cyberattacks and infrastructure operators. It is the same model, but with protections removed in certain areas. According to the company, Mythos 5 has the strongest cybersecurity capabilities among all models in the world. The names were chosen deliberately: Fable comes from the Latin fabula, “that which is told,” related to the Greek mythos, and the models are distinguished precisely by their protections.
Coding, analytics, and vision
According to Anthropic, both models work autonomously longer than previous Claude versions. In software development, Stripe reported during early testing that Fable 5 compressed months of engineering work into days: in a 50-million-line Ruby codebase, the model completed a codebase-wide migration in one day that would have taken a team more than two months manually. On Cognition’s FrontierCode evaluation, Fable 5 achieved the highest score among frontier models even at a medium effort level.
In analytical tasks, Fable 5 showed the best result on the Hebbia financial benchmark, while trading company IMC noted that it passed almost all of its analysis tests. In computer vision tasks, the model became a new benchmark: it extracts precise numbers from scientific charts and reconstructs a web application’s source code from screenshots alone. As an illustration, Fable 5 completed Pokémon FireRed using only screen frames, without maps or hints, whereas earlier models required complex auxiliary scaffolding. In terms of memory, the model retains context over millions of tokens and improves its result using its own notes: in the turn-based game Slay the Spire, access to persistent file memory boosted its performance three times more than it did for Opus 4.8.
Drugs, hypotheses, and genomics
Anthropic links the main scientific results to Mythos 5. According to the company, its protein design specialists accelerated part of the drug development process by about 10 times. Mythos 5, using tools for protein design and bioinformatics, matched or surpassed experienced operators without human involvement: the model selected binding sites itself, launched tools, and corrected its own errors. Out of 14 protein targets, 9 produced strong candidates for further development.
Mythos 5 became the first Anthropic model to consistently produce new scientific hypotheses. In a blind comparison with Opus-class models, the company’s scientists preferred Mythos hypotheses in molecular biology in about 80% of cases. One of them, a new mechanism of action for an Escherichia coli (E. coli) protein, was independently confirmed by another laboratory. In addition, over a week of mostly autonomous work, the model conducted a genomics study: it collected data on millions of cells from 138 animal species and trained its own model, which, according to Anthropic, outperformed a recent development published in Science while being 100 times smaller.
Classifiers and jailbreak protection
According to Anthropic, Mythos-class models have reached a threshold beyond which serious risks emerge. To release Fable 5 safely, the company added a set of classifiers: separate systems that detect potential misuse, including attempts to bypass protections, and prevent the main model from responding. They cover three areas: cybersecurity, biology and chemistry, and “distillation” attempts to extract the model’s capabilities to train competing systems. When triggered, the request is handled by Opus 4.8, and more than 95% of Fable sessions proceed without such switching.
The company tested the protections for resistance to “jailbreaks,” meaning attempts to bypass restrictions. An external vulnerability discovery program found no universal bypass after more than 1,000 hours of testing, although the UK’s AI Security Institute (AISI) made progress in that direction. In one of the tests, Fable 5 did not complete a single harmful request involving cyberattack planning, exploit development, or protection bypassing, including when 30 public jailbreak techniques were used.
The biology and chemistry protections are configured especially broadly. In a test on assembling the capsid of adeno-associated virus (AAV), a component used for gene therapy delivery, Mythos-class models surpassed specialized protein models through biological reasoning. The same capability in the wrong hands could help design dangerous viruses, so for now Fable 5 routes most biology and chemistry requests to Opus 4.8.
At the same time, in an automated alignment evaluation, the level of undesirable behavior in Mythos 5, including deception and assistance with misuse, was low and comparable to Opus 4.8. Since Fable 5 is the same model, its alignment is the same. The company provides details in the model’s system card.
Pricing and access to the new models
For Fable 5, Mythos 5, and future models of the same level, Anthropic is introducing mandatory 30-day data retention for business customers. The company promises not to use this data for training and to delete it after 30 days, while logging access to it. It explains this by saying that the data helps defend against sophisticated attacks and reduce false positives.
Both models cost $10 per million input tokens and $50 per million output tokens, which is less than half the price of Mythos Preview. Fable 5 is available everywhere starting today, including through the Claude API. Mythos 5 is currently limited to Glasswing partners. For subscriptions, Anthropic is rolling out access in stages: from June 9 to June 22, Fable 5 is included in Pro, Max, Team, and Enterprise seat-based plans at no extra charge, and starting June 23 it will require token-based payment.
