Anthropic Unleashes Claude Fable 5, Mythos 5

Anthropic has unveiled Claude Fable 5, its most capable AI model to date, designed for broad public access. Alongside it, the company is releasing Claude Mythos 5, a version with enhanced cybersecurity capabilities for a select group.

Fable 5 sets new benchmarks, excelling in software engineering, knowledge work, vision, and scientific research. Its performance scales with task complexity, surpassing previous Anthropic models on longer, more intricate assignments. This launch, detailed by Anthropic News, marks a significant leap in AI capability.

Recognizing the potential risks associated with such power, Fable 5 incorporates new safeguards. Queries touching on sensitive areas like cybersecurity are rerouted to the next-tier model, Claude Opus 4.8. These safeguards are conservatively tuned, leading to occasional false positives, but aim for minimal disruption, with less than 5% of sessions affected on average.

Claude Mythos 5, built on the same foundation as Fable 5 but with some safeguards lifted, is initially deployed through Project Glasswing in collaboration with the US government. This initiative, an upgrade to Claude Mythos Preview, focuses on advanced cybersecurity applications, building on Anthropic's previous work with Project Glasswing.

Unprecedented Performance

Fable 5's prowess is evident in real-world applications. Stripe reported that Fable 5 condensed months of engineering work into days, executing a codebase-wide migration on a 50-million-line Ruby codebase in a single day.

On Cognition's FrontierCode evaluation, Fable 5 achieved the highest score among leading models for its ability to produce high-quality production code.

In financial analysis, Fable 5 leads on Hebbia's Finance Benchmark, demonstrating substantial gains in document interpretation and complex problem-solving.

IMC noted Fable 5's near-perfect performance on trading-analysis evaluations, covering factual lookup, reasoning, and root-cause analysis.

The model also redefines state-of-the-art in vision tasks, capable of extracting precise data from complex scientific figures and reconstructing web app source code from screenshots alone.

Fable 5 notably completed the game Pokémon FireRed using only visual input, a task that previously required extensive AI assistance.

Long-context capabilities allow Fable 5 to maintain focus over millions of tokens, improving outputs via self-generated notes.

Performance in games like Slay the Spire saw a threefold improvement with persistent memory access compared to Opus 4.8.

Fable 5 autonomously simulated solar system dynamics and predicted eclipses from first principles.

It autonomously played Factorio, strategizing and building complex factories.

Fable 5 designed a complete 3D-printable model and the CAD editor itself, including its AI copilot.

A fluid simulation, synchronized to a classical EDM remix also produced by Fable 5, showcased its creative coding abilities.

Mythos 5: Cybersecurity and Research Prowess

For specialized applications, Mythos 5 accelerates drug design, with internal experts seeing a tenfold speedup in certain aspects.

Mythos 5 matched or exceeded skilled human operators in protein design tasks, operating autonomously with bioinformatics tools.

The model generated novel, compelling hypotheses in molecular biology, preferred by scientists over Opus-class models 80% of the time.

One Mythos 5 hypothesis regarding an E. coli protein mechanism was independently corroborated.

In genomics, Mythos 5 conducted autonomous research, analyzing single-cell data across 138 species and training a custom ML model that outperformed a recent Science publication's model.

The trained genomics model from Mythos 5 was 100 times smaller than the benchmarked model.

Automated alignment assessments show Mythos 5 has low levels of misaligned behavior, comparable to Opus 4.8.

Pricing and Accessibility

Claude Fable 5 and Mythos 5 are priced at $10 per million input tokens and $50 per million output tokens.

This pricing is less than half that of Claude Mythos Preview, reflecting Anthropic's goal of broader AI access.

The release underscores Anthropic's commitment to making advanced AI capabilities accessible quickly and safely.

New safety classifiers are employed to detect and prevent misuse, particularly in cybersecurity and biology.

When classifiers trigger, Fable 5 responses are handled by Claude Opus 4.8, ensuring a capable fallback.

Over 95% of Fable 5 sessions operate without any fallback, maintaining peak performance.

Safeguards cover cybersecurity, biology, chemistry, and distillation to prevent malicious use.