Claude Fable 5 vs Claude Mythos 5: Benchmarks, Safeguards, and Pricing Explained

Software Engineering

The most striking real-world data point comes from Stripe. In a 50-million-line Ruby codebase, Fable 5 completed a codebase-wide migration in a single day — a task that would have required a full engineering team more than two months by hand. On Cognition’s FrontierCode evaluation, which measures whether models can pass difficult coding tasks while meeting production-quality standards, Fable 5 scores highest among frontier models even at medium effort.

Cursor’s CEO Michael Truell noted that Fable 5 is the state-of-the-art model on CursorBench and has

opened up a class of long-horizon problems that were out of reach for earlier models.

GitHub’s Chief Product Officer echoed this, citing autonomy and reliability on complex, long-horizon coding tasks as the defining improvement over previous generations.

Knowledge Work and Finance

On Hebbia’s Finance Benchmark for senior-level reasoning, Fable 5 achieves the highest score of any model tested. Gains are particularly pronounced in document-based reasoning, chart and table interpretation, and problem-solving under ambiguity.

IMC’s trading-analysis evaluations showed Fable 5 performing near-perfectly across factual lookup, conceptual reasoning, root-cause analysis, and expected-value analysis. These are not toy benchmarks — they reflect the kind of multi-step analytical work that defines high-stakes knowledge work environments.

Vision

Fable 5 sets a new standard for vision-based tasks. It can extract precise numerical data from complex scientific figures and reconstruct a web application’s source code from screenshots alone — without additional scaffolding. The Pokémon FireRed demonstration is illustrative: earlier Claude models required a complex helper harness to navigate the game, while Fable 5 completed it using only raw game screenshots and no supplementary tools.

Memory and Long-Context Reasoning

Fable 5 maintains coherence across millions of tokens in extended tasks and actively improves its outputs using persistent, file-based memory. In Slay the Spire testing, providing Fable 5 with persistent memory improved performance three times more than the same intervention did for Opus 4.8. The model also reached the game’s final act three times more frequently — a proxy for sustained strategic reasoning over long task horizons.

Cybersecurity

Anthropic states plainly that Mythos 5 has the strongest cybersecurity capabilities of any model in the world. This includes not only vulnerability discovery and exploitation but also agentic hacking — the ability to execute multiple phases of a cyberattack autonomously, including reconnaissance, lateral movement, and recovery from failures. For defensive applications, this capability is enormously valuable. For offensive misuse, it would be equally dangerous, which is precisely why it remains behind a trusted-access gate.

Drug Design and Life Sciences

Internal protein design experts using Mythos 5 accelerated aspects of the drug design process by approximately ten times. In one structured evaluation, Mythos 5 — operating with protein design and bioinformatics tools but no human assistance — matched or exceeded skilled human operators across the full scientific workflow: selecting binding sites, running design tools, and recovering from intermediate failures. Nine of fourteen protein targets yielded strong drug design candidates currently under investigation.

Novel Scientific Hypotheses

Mythos 5 is Anthropic’s first model to consistently generate novel, compelling scientific hypotheses. In blinded comparisons against Opus-class models, Anthropic’s scientists preferred Mythos 5’s molecular biology hypotheses 80% of the time. One hypothesis — a novel mechanism for an E. coli protein — was independently corroborated by a separate laboratory working on the same problem.

Autonomous Genomics Research

In a week-long largely autonomous research task, Mythos 5 assembled single-cell data for millions of cells across 138 animal species, then designed and trained a custom machine learning model to identify functionally equivalent cells across distantly related organisms. The resulting model outperformed a recently published model from the journal Science — while being 100 times smaller. Anthropic intends to publish these results.

Classifier Coverage

Fable 5 runs a separate layer of AI classifiers that monitor incoming requests in real time. When a request is flagged — for cybersecurity, biology and chemistry, or distillation topics — the response is automatically routed to Claude Opus 4.8 instead. Users are informed when this occurs. The fallback is not a refusal; it is a response from a highly capable model, which Anthropic frames as a meaningfully better user experience than a hard block.

Anthropic’s own data shows that more than 95% of Fable 5 sessions involve no fallback at all. For those sessions, Fable 5’s performance is effectively equivalent to Mythos 5.

Deliberate Conservatism and False Positives

Anthropic acknowledges that the classifiers are intentionally tuned conservatively. Some benign requests will trigger them. This is a calculated trade-off: the company has prioritized minimizing the risk of harmful uplift over minimizing user friction at launch. The stated goal is to reduce false positives iteratively as the safeguards are refined post-launch.

This transparency is notable. Rather than presenting the safeguards as a solved problem, Anthropic is treating them as a living system — one that will improve as more capable models arrive and as classifier precision increases.

Alignment Assessment

Automated alignment testing found that Mythos 5’s level of misaligned behavior — including deception and cooperation with user misuse — was low and comparable to Opus 4.8. Since Fable 5 is the same underlying model, its alignment profile is expected to be similar. Full details are documented in the model’s system card.