Claude Mythos 5 Unleashed: 50 Million Lines of Code Processed in One Day

Model variants

Anthropic released two versions of its flagship model: Claude Fable 5 , which includes a safety “net” and is publicly accessible, and Claude Mythos 5 , the full‑power “blood‑only” version limited to trusted users. Both share the same underlying architecture, so their core technical specifications are identical.

Safety mechanism and model routing

Fable 5 runs a set of independent classifiers that detect high‑risk requests in three domains: cybersecurity, bio‑chemical threats, and model‑distillation attacks. When a request triggers a classifier, the system automatically routes the query to the previous‑generation model Claude Opus 4.8, providing a degraded but still useful answer instead of a blunt refusal. Anthropic reports that over 95 % of Fable 5 conversations do not trigger degradation; the remaining <5 % are subject to stricter safety pathways, and the classifiers are deliberately conservative, leading to occasional false positives on legitimate research or security‑testing tasks.

Benchmark performance

Software engineering (SWE‑bench Pro) : Claude Fable 5 achieved 80.3 % accuracy, compared with GPT‑5.5’s 58.6 %.

Frontier Code (Cognition) : In the medium‑effort setting, Fable 5 obtained the highest score among frontier models.

Real‑world code migration : A 50 million‑line Ruby codebase at Stripe was fully migrated by Fable 5 in one day, a task that would normally require over two months of engineering effort.

ViBench (frontend development) : Fable 5 demonstrated near‑saturation on basic development use‑cases, achieving true one‑shot application generation.

Visual reasoning (GDPpdf) : Fable 5 and Mythos 5 scored 29.8 % without external tools, surpassing Opus 4.8 (22.5 %), GPT‑5.5 (24.9 %) and Gemini 3.1 Pro (16.7 %). A demo showed the model playing the Pokémon game entirely from raw screen captures.

Long‑context and memory : Anthropic states that Fable 5 can maintain focus over million‑token tasks and improve its output using persistent notes. In a Slay‑the‑Spire test, adding file‑level memory boosted performance threefold over Opus 4.8 and tripled the probability of reaching the final chapter.

Finance and analytics (Hebbia benchmark) : Fable 5 broke the 90 % threshold, delivering double‑digit gains on long‑document reasoning, chart interpretation, and multi‑step root‑cause analysis. In quantitative‑trading evaluations by IMC and Optiver, the model achieved near‑perfect stability across repeated runs, scoring full weight on fact‑retrieval, conceptual reasoning, and expected‑value calculation.

Frontier physics (VibeCAD) : Using only one‑third of the inference tokens, Fable 5 produced physics research results in 36 hours that required four days on GPT‑5.5.

Protein design (Mythos 5) : In a bio‑medical workflow without human assistance, Mythos 5 designed 14 protein‑target complexes, of which nine entered real drug‑development pipelines. The model also generated novel scientific hypotheses that were independently validated in peer‑reviewed studies.

Token efficiency and pricing

Both models are priced at $10 per million input tokens and $50 per million output tokens. Anthropic emphasizes token efficiency to keep the cost of long‑running autonomous tasks manageable.

Data retention policy

All traffic to Mythos‑level models is retained for 30 days for security monitoring (e.g., detecting complex attacks, new jailbreaks, and cross‑request exploits). The retained data are not used for model training.

Agent capabilities and autonomous demonstrations

Visual game playing : Without any external scaffolding, Fable 5 processed raw Pokémon screen captures and completed the entire game autonomously.

Slay‑the‑Spire : When equipped with persistent file‑level memory, the model’s success rate increased threefold.

Mollick internal test : Professor Ethan Mollick fed a 15‑page project design to Fable 5, providing only high‑level requirements. Over roughly nine hours, the model autonomously spawned multiple agents to research, outline, code, and verify the solution, delivering a high‑quality product without human intervention.

Implications for AI collaboration

The combination of long‑context reasoning, autonomous agents, and safety‑aware routing signals a shift from “wizard‑style” prompting—where users iteratively steer the model—to a “patron” relationship, where the model operates as an autonomous studio and humans act as overseers who validate final deliverables.

Code example

[1]https://www.anthropic.com/news/claude-fable-5-mythos-5
[2]https://www.oneusefulthing.org/p/what-it-feels-like-to-work-with-mythos
[3]https://www.biorxiv.org/content/10.64898/2026.03.12.711259v1