Anthropic released two new models built on the same foundation: Claude Fable 5, available to all users and developers starting immediately, and Claude Mythos 5, a version with certain safety restrictions lifted that is currently accessible only to a small group of vetted partners. Both are priced at $10 per million input tokens and $50 per million output tokens – less than half the cost of Claude Mythos Preview, which the models replace.
Fable 5 is Anthropic’s most capable model to date and the first Mythos-class system made available for general use. The company describes it as state-of-the-art across nearly all tested capability benchmarks, with performance advantages that grow as tasks become longer and more complex. Mythos 5 is the same model with safeguards lifted in specific areas where Anthropic determined the risks were too high for unrestricted public access.
What the Models Can Do
Software engineering
- Stripe, a payments company with early access, compressed multiple months of engineering work into days using Fable 5
- In a 50-million-line Ruby codebase, the model completed a full migration in one day – work Stripe estimated would take a full human team more than two months
- On Cognition’s FrontierCode evaluation, which tests production-quality code output, Fable 5 scores highest among frontier models even at medium effort
Knowledge work
- Highest score on Hebbia’s Finance Benchmark for senior-level analytical reasoning, with gains in document analysis, chart interpretation, and problem solving
- Trading firm IMC reported near-complete performance on their internal trading-analysis evaluations, spanning factual lookup, conceptual reasoning, root-cause analysis, and expected-value work
Vision
- Extracts precise data from complex scientific figures
- Can reconstruct a web application’s source code from screenshots alone
- Completed the video game Pokemon FireRed using only raw visual input with no supplementary tools – earlier Claude models required additional scaffolding to attempt the same task
Memory and long-context
- Maintains focus across millions of tokens in extended sessions
- In testing with the deck-building game Slay the Spire, persistent file-based memory improved Fable 5’s performance three times more than it improved Opus 4.8’s; Fable also reached the game’s final act three times as often
Drug design and life sciences (Mythos 5)
- Internal protein design experts using Mythos 5 accelerated parts of the drug design process by around ten times
- The model autonomously selected binding sites, chose and ran protein design tools, and recovered from failures without human assistance – matching or exceeding skilled human operators
- Nine of fourteen protein targets in one study produced viable drug design candidates currently under investigation
- Scientists preferred Mythos 5’s molecular biology hypotheses roughly 80% of the time in blind comparisons against earlier models; one hypothesis about an E. coli protein mechanism was independently corroborated by a separate research lab
- In over a week of largely autonomous work, Mythos 5 assembled single-cell data across 138 animal species, trained a custom machine learning model to identify functionally equivalent cells across distantly related organisms, and outperformed a recently published Science journal model despite being 100 times smaller
The Safety Architecture
Releasing a model with Mythos-class capabilities to the public required new safeguards. Anthropic built a layer of separate AI classifiers that intercept potentially harmful requests before Fable 5 responds. When triggered, those requests are automatically rerouted to Claude Opus 4.8. Anthropic says this occurs in fewer than 5% of sessions, and that users are notified each time a fallback happens.
Three categories are covered:
Cybersecurity – Fable 5 will not assist with exploit development, offensive hacking, or cyberattack planning. In testing, the model complied with zero harmful single-turn requests across these categories, including against 30 different publicly known jailbreak techniques. An external bug bounty program produced no universal jailbreaks after more than 1,000 hours of testing. The UK AI Security Institute made progress toward one during an initial testing window, which Anthropic acknowledged – the company’s stated goal is not to eliminate jailbreaks entirely but to make them slow and costly enough to detect before they spread.
Biology and chemistry – Given the dual-use risk of advanced biological reasoning – the same capability that aids gene therapy research could assist in designing harmful agents – Fable 5 falls back to Opus 4.8 on most biology and chemistry queries. Anthropic cited its own testing in which Mythos-class models outperformed dedicated protein modeling tools on predicting properties of adeno-associated viruses, illustrating how general scientific reasoning can cross into sensitive territory without explicit training. The company says these safeguards are deliberately broad and will be narrowed as trusted access programs expand.
Distillation – Requests identified as attempts to systematically extract Fable 5’s capabilities for use in training competing models – a practice Anthropic says it has detected from actors in authoritarian countries – are blocked and rerouted.
Trusted Access and Mythos 5
Claude Mythos 5 launches today as an upgrade for all current Claude Mythos Preview users, primarily the cybersecurity organizations and critical infrastructure providers enrolled in Project Glasswing. In this context the model operates with cybersecurity safeguards removed.
A separate trusted access program for biology is opening shortly. It will allow a limited group of life science researchers to use Fable 5 with biology and chemistry safeguards lifted while cyber restrictions remain in place. Anthropic says it plans to expand both programs over time and to publish the genomics research conducted by Mythos 5.
Data Retention and Pricing
A new data policy applies to all Mythos-class model traffic. Anthropic will retain interaction data for 30 days across both consumer and enterprise surfaces – not for model training but for safety monitoring, jailbreak detection, and false positive reduction. All human access to the data is logged, and the data is deleted after 30 days in almost all cases.
Pricing: $10 per million input tokens, $50 per million output tokens for both models. Available via the Claude API at model string claude-fable-5.
Subscription plan rollout:
- Available at no extra cost on Pro, Max, Team, and seat-based Enterprise plans from today through June 22.
- From June 23, usage credits will be required on subscription plans.
- Anthropic plans to restore Fable 5 as a standard subscription tier once capacity allows, with advance notice before any changes.
