Anthropic on Tuesday released Claude Fable 5, identifying it as a Mythos-class model that the company has prepared for broad use. Anthropic described Fable 5 as the most capable model it has made available to the public to date, reporting state-of-the-art results across nearly all capability benchmarks the company tested.
According to Anthropic, Fable 5 demonstrates strong, cross-domain performance in software engineering, knowledge work, vision, scientific research, and related areas. The company highlighted that the models performance advantage over its earlier offerings grows with task length and complexity.
As part of the public release, Anthropic implemented new safeguards designed to intercept certain categories of queries and route them to Claude Opus 4.8, which the company describes as its next-most-capable model. Anthropic said these safeguards activate in fewer than 5% of sessions on average, though it acknowledged that the classifiers sometimes flag harmless requests. The safeguards were tuned conservatively, Anthropic said, to enable a safe and rapid rollout of Fable 5.
In parallel, Anthropic launched Claude Mythos 5 for a small cohort of cyberdefenders and infrastructure providers. Mythos 5 runs on the same underlying model as Fable 5 but has restrictions relaxed in certain areas. The initial deployment for Mythos 5 will be conducted through Project Glasswing in partnership with the U.S. Government as an upgrade to Claude Mythos Preview. Anthropic plans to broaden access to Mythos 5 over time through a trusted access program.
Both Fable 5 and Mythos 5 are priced at $10 per million input tokens and $50 per million output tokens, a rate Anthropic said is less than half the cost of Claude Mythos Preview.
Early usage and benchmark reports accompanying the release show notable applications and evaluations. In internal or partner testing, Stripe reported that Fable 5 completed a codebase-wide migration in a 50-million-line Ruby project in one day - a task Anthropic said would have required a team more than two months to finish manually. On Cognitions FrontierCode evaluation, Fable 5 achieved the top score among frontier models.
On domain-specific benchmarks, Anthropic noted that Fable 5 scored highest on Hebbias Finance Benchmark for senior-level reasoning among models tested. IMC provided additional trading-analysis assessments, in which Fable 5 performed well on factual lookup, conceptual reasoning, root-cause analysis, and expected-value analysis.
Anthropic also emphasized the models vision and multimodal capabilities. The company reported that Fable 5 can extract numerical data from scientific figures and carry out complex vision-based tasks such as reconstructing a web applications source code from screenshots. In a demonstration of gaming-related vision capability, Fable 5 reportedly completed the game Pokémon FireRed using only raw game screenshots, without access to maps, navigation aids, or supplemental game-state information.
In areas touching biomedical research, Anthropic reported results from Mythos 5 and internal teams. Protein design experts at Anthropic stated that Mythos 5 accelerated certain drug design tasks by about ten times, matching or exceeding skilled human operators when using protein design and bioinformatics tools without human assistance. Of 14 protein targets evaluated in this study, nine yielded strong candidate molecules now under investigation.
Anthropic further described a largely autonomous genomics research effort conducted by Mythos 5 that ran for more than a week. In this work the model assembled single-cell data covering millions of cells across 138 animal species, and then designed and trained a machine learning model. With only high-level human guidance, Anthropic said the trained model outperformed a recent model published in the journal Science while being 100 times smaller. The company stated it intends to publish these results in the coming months.
To prevent misuse and help guard against jailbreaks, Anthropic added classifiers and routing rules to the Fable 5 deployment. When Fables classifiers detect requests related to cybersecurity, biology and chemistry, or distillation, the request is handled by Claude Opus 4.8 instead, and users are informed when such rerouting occurs.
Anthropic reported results from external hardening efforts. A bug bounty program and external red-teaming produced no universal jailbreaks after more than 1,000 hours of testing. Outside red-teaming organizations also did not identify universal jailbreaks on long-form agentic tasks, though Anthropic said the UK AISI made progress toward such a jailbreak within a brief initial testing window.
One unnamed external partner observed that Fable 5s safeguards against harmful cyber queries were the strongest of any model it tested, including Opus 4.8 and Opus 4.7. Anthropic noted that Fable 5 complied with zero harmful single-turn requests relating to planning a cyberattack, exploit development, or defense evasion in the evaluations cited.
Anthropic introduced a new data retention policy tied to Mythos-class models, requiring 30-day retention for all traffic on these systems. The company said it will not use that retained data to train new Claude models or for any non-safety-related purpose, and that the data will be deleted after 30 days in almost all cases.
Regarding access, Anthropic said all users who currently have access to Claude Mythos Preview will be able to upgrade to Claude Mythos 5. The company plans to expand access through a trusted access program that allows cybersecurity organizations to apply systematically, and it intends to open a separate trusted access program for biology to help accelerate biomedical research.
Availability and plan details were provided alongside the launches. Claude Fable 5 is available everywhere starting Tuesday. Claude Mythos 5 is restricted to Glasswing partners until the trusted access program becomes operational.
On the Claude API and consumption-based Enterprise plans, Anthropic said Fable 5 is fully available starting Tuesday. For subscription plans, Fable 5 will be included on Pro, Max, Team, and seat-based Enterprise plans at no additional cost from Tuesday through June 22. Beginning June 23, using Fable 5 will require usage credits; Anthropic said it aims to reinstate Fable 5 as a standard subscription offering when sufficient capacity permits.
Key points
- Anthropic released Claude Fable 5, a Mythos-class model that it says represents its most capable public offering, and also launched Claude Mythos 5 for a limited set of trusted partners.
- New safety measures route certain cybersecurity, biology, chemistry, and distillation queries to Claude Opus 4.8, with these safeguards triggering in under 5% of sessions on average; Anthropic tuned the system conservatively.
- Anthropic reported strong benchmark and application performance in software engineering, finance reasoning, vision tasks, and drug design, and set a 30-day data retention policy for Mythos-class traffic.
Risks and uncertainties
- Safeguards sometimes catch harmless requests - the conservative tuning that enabled a rapid release means legitimate queries may be routed away from Fable 5, potentially affecting workflows in software development, finance, or research.
- External red-teaming did not find universal jailbreaks during testing, but progress toward a jailbreak was reported in an initial window by one organization, indicating an ongoing security risk for cybersecurity and infrastructure users.
- Access to Mythos 5 remains limited to Glasswing partners until the trusted access program expands, creating uncertainty for organizations in cybersecurity and biology that seek immediate access.