Claude Fable 5 and Mythos 5: A Complete Guide to Anthropic’s New AI Models

On June 9, 2026, Anthropic did something it had explicitly said it was not planning to do: it gave the public access to a model from the same capability tier as Claude Mythos, the AI system the company had described in April as too risky for general release. The model that arrived is called Claude Fable 5, and it is, in Anthropic’s own words, the most capable model the company has ever made available. Alongside it, a small group of cybersecurity partners and infrastructure providers received an upgrade to a second model, Claude Mythos 5, which is built on the same underlying system but with key safety restrictions lifted.

The two-model release is unusual, the naming is deliberate, and the story behind both says a great deal about how Anthropic is now navigating the gap between AI capability and AI safety in real time.

What Are Claude Fable 5 and Claude Mythos 5?

Fable 5 and Mythos 5 are, technically, the same model. Anthropic has been explicit about this: both are built on identical underlying weights, the same training run, the same core capabilities. What differs between them is the safety layer wrapped around that core. Mythos-class is the name Anthropic has given to a tier of models that sit above its Opus class in capability, the first of which, Claude Mythos Preview, was released in a deliberately restricted form in April 2026 through Project Glasswing, Anthropic’s collaboration with the US government on AI-assisted cybersecurity.

Fable 5 and Mythos 5 are the second generation of this Mythos-class tier, and the first to be offered as a pair: one designed for general use, one designed for users who need capabilities the general version deliberately withholds.

• Claude Fable 5: State-of-the-art on nearly all tested benchmarks of AI capability. Exceptional performance in software engineering, knowledge work, vision, and scientific research. Ships with safety classifiers that automatically reroute high-risk queries, in areas like cybersecurity and biology, to Claude Opus 4.8 instead. More than 95% of sessions never trigger a fallback at all.
• Claude Mythos 5: The same underlying model as Fable 5, with cybersecurity safeguards lifted for a small group of cyber defenders and critical infrastructure providers. Deployed initially through Project Glasswing in collaboration with the US government. Described by Anthropic as having the strongest cybersecurity capabilities of any model in the world.

The name “Fable” is not incidental. Anthropic notes that Fable derives from the Latin fabula, “that which is told”, a word closely related to the Greek mythos. The two names describe the same underlying story, told two different ways: one version told to everyone, one version told only to those who need the unabridged edition.

Why Anthropic Created Two Versions of the Same Frontier AI

The decision to split one model into two products is a direct response to a problem that has become increasingly central to frontier AI development: dual-use capability. A model that is extremely good at finding and exploiting software vulnerabilities is simultaneously one of the most valuable tools available to a defensive cybersecurity team and one of the most dangerous tools available to an attacker. A model that can reason about viral protein structures with the sophistication needed to accelerate gene therapy research is also a model that could, in principle, assist someone attempting to design a dangerous pathogen.

Anthropic’s own testing illustrates the scale of this capability: when Mythos-class models were evaluated on a task related to predicting how genetic modifications affect the assembly of adeno-associated virus shells, a component used in gene therapy delivery, they outperformed specialised protein-language models without having been explicitly trained for the task.

Faced with a model this capable, Anthropic’s options were limited to three. Release it to everyone with no restrictions, accepting the uplift risk to malicious actors. Restrict it to a small group indefinitely, limiting the benefits the technology could provide to the much larger population of legitimate users, including the cybersecurity defenders, researchers, and developers who could use it productively. Or build a safety layer robust enough to separate the two populations, releasing the capability broadly while keeping its most dangerous applications gated. Anthropic chose the third option, and Fable 5 and Mythos 5 are the result.

As the company put it in its announcement, this is part of what it calls a “race to the top”: demonstrating that frontier capability and meaningful safety guardrails are not mutually exclusive, and that the industry can compete on both simultaneously rather than trading one for the other.

Claude Fable 5: The Public Face of Anthropic’s Most Advanced AI

Fable 5’s headline claim is unambiguous: its capabilities exceed those of any model Anthropic has ever made generally available, and the longer and more complex the task, the larger its lead over previous Claude models becomes. The early customer feedback Anthropic published alongside the launch is unusually specific and comes from companies with direct, hands-on testing experience rather than marketing partnerships.

Stripe reported that Fable 5 compressed months of engineering into days, performing a codebase-wide migration across a 50-million-line Ruby codebase in a single day, a task that would otherwise have required a full team more than two months of manual work. On Cognition’s FrontierCode evaluation, which tests whether models can complete difficult coding tasks to production-quality standards, Fable 5 scored highest among frontier models, including at medium reasoning effort, while also being more token-efficient than previous Claude models.

Cursor’s CEO Michael Truell described Fable 5 as state of the art on CursorBench, opening up a class of long-horizon problems that were previously out of reach. GitHub’s Chief Product Officer Mario Rodriguez said the model handled complex, long-horizon coding tasks with a level of autonomy and reliability exceeding previous benchmarks.

• Software Engineering (A 50-million-line migration, done in a day): Stripe’s early testing found Fable 5 completed a codebase-wide migration across a 50-million-line Ruby codebase in a single day, a task that would have taken a full engineering team over two months by hand.
• Knowledge Work (Top score on finance and legal reasoning): 
• Highest score of any model on Hebbia’s Finance Benchmark for senior-level reasoning. IMC said Fable 5 aced its trading-analysis evaluations nearly across the board, including root-cause and expected-value analysis.
• Vision (Rebuilds source code from screenshots alone): New state of the art on vision tasks. Can extract precise numbers from scientific figures and reconstruct a web app’s source code purely from screenshots. Beat Pokémon FireRed using vision alone, with no helper harness, previous Claude models needed one.
• Memory and Long-Context (Three times the performance with persistent memory): When given access to persistent file-based memory while playing Slay the Spire, Fable 5’s performance improved three times more than Opus 4.8’s did under the same conditions, and it reached the game’s final act three times more often.

Beyond benchmarks, Anthropic demonstrated Fable 5’s autonomous capabilities through a series of open-ended tasks: building a simulation of the solar system from physics first principles to predict solar eclipses, playing the factory-building game Factorio by independently strategising and constructing an automated production line, designing a complete 3D-printable model inside a browser-based CAD editor that Fable 5 itself had built, including the AI copilot inside that editor, and producing a fluid simulation synchronised to the beat of an EDM remix that the model composed itself, having never previously “heard” music.

Claude Mythos 5: The Restricted Model Built for High-Stakes Research

Where Fable 5 is defined by its safeguards, Mythos 5 is defined by their absence in specific domains. Anthropic describes Mythos 5 as having the strongest cybersecurity capabilities of any model in the world, and the use cases it has already enabled are correspondingly specialised. In drug design, Anthropic’s internal protein design experts used Mythos 5 to accelerate aspects of the process by around ten times.

In one documented example, the model, given protein design and bioinformatics tools but no human assistance, matched or beat skilled human operators across the full workflow: choosing binding sites, selecting and running design tools, and recovering from failures along the way. Nine of fourteen protein targets in this study yielded strong drug design candidates that Anthropic is now investigating further.

In molecular biology, Mythos 5 is described as Anthropic’s first model to consistently produce novel, compelling scientific hypotheses. In blinded comparisons, Anthropic’s scientists preferred Mythos 5’s hypotheses over those from Opus-class models roughly 80 percent of the time, and several have been advanced to experimental testing. One hypothesis, concerning a novel mechanism for an E. coli protein, was independently corroborated by a separate laboratory working on the same problem.

In genomics, Mythos 5 conducted what Anthropic calls novel research over more than a week of largely autonomous work: it assembled single-cell data spanning 138 animal species, designed and trained a custom machine learning model to identify functionally equivalent cells across distantly related species, and with only high-level human guidance, produced a model that outperformed a recently published Science paper despite being one hundred times smaller.

The Biggest Difference Between Fable 5 and Mythos 5

The single most important fact about the relationship between these two models is this: they are the same model. Anthropic has stated this directly, and the implication is significant. The difference between Fable 5 and Mythos 5 is not a difference in intelligence, training, or underlying capability. It is a difference in what each model is allowed to act on. Fable 5 ships with safety classifiers, separate AI systems that monitor requests and detect potential misuse, that automatically intercept queries related to three categories and reroute them to the less capable Claude Opus 4.8 instead.

• Cybersecurity: Mythos-class models excel at discovering and exploiting software vulnerabilities, including agentic hacking that combines reconnaissance, exploit discovery, and lateral movement. Fable 5’s classifiers are designed to prevent any progress on offensive cyber tasks, falling back to Opus 4.8 instead. One external partner found Fable 5’s safeguards against harmful cyber queries to be the most robust of any model tested.
• Biology and Chemistry: Beyond the narrow bioweapons-related queries Anthropic has long blocked, Fable 5’s classifiers now cover a broader range of biology and chemistry requests, given the model’s demonstrated ability to complete real-world scientific tasks such as predicting viral shell assembly properties relevant to gene therapy and, in principle, dangerous pathogen design.
• Distillation: Anthropic has previously identified large-scale attempts to “distill” Claude’s capabilities into competing models, including in jurisdictions without comparable safety standards. Requests flagged as part of such attempts are intercepted by Fable 5’s classifiers and rerouted to Opus 4.8, preventing proliferation of near-frontier capability without appropriate safeguards.

Importantly, Anthropic reports that these classifiers trigger on average in less than 5 percent of sessions, and that the safeguards have been deliberately tuned conservatively, meaning they will sometimes catch entirely harmless requests. The company has framed this trade-off explicitly: it would rather over-block in the early period after launch and refine the classifiers to reduce false positives over time, than under-block and create a meaningful uplift risk on day one.

In the external red-teaming Anthropic commissioned, more than 1,000 hours of bug bounty testing produced no universal jailbreak, defined as a prompt, script, or method that allows a user to interact with the model as though its safeguards were not present, though the UK’s AI Safety Institute reportedly made progress toward one within a brief initial testing window.

What Claude Fable 5 and Mythos 5 Tell Us About the Future of AI

The timeline from Mythos Preview to Fable 5 and Mythos 5 is, on its own, one of the more revealing data points about the current pace of frontier AI development. In April 2026, Anthropic released Claude Mythos Preview to a deliberately tiny group of partners, citing concerns serious enough that the company explicitly said it did not plan to make the model generally available. Less than three months later, a version of that same model is available to anyone with a Claude account. What changed in the interim was not the model. It was the safeguards around it, developed and validated quickly enough that Anthropic judged the combination of capability and protection ready for broad release.

This compression has two readings, and both are probably true simultaneously. The first is that Anthropic’s safety research is genuinely advancing at a pace that allows previously withheld capabilities to be released responsibly within months rather than years, which is a meaningfully positive signal for an industry whose critics have long argued that safety work cannot keep pace with capability work.

The second is that the commercial and competitive pressure to deploy frontier capability is intense enough that a three-month gap between “too dangerous for general release” and “available to everyone” represents the practical floor of how long any AI lab can sit on a finished, highly capable model before market pressure, competitor releases, and customer demand make withholding it untenable.

Anthropic’s own framing, that this is part of a “race to the top” in which the industry competes on building the strongest safeguards rather than the weakest ones, is the optimistic version of this story. Whether that framing holds will depend less on this specific release and more on how the industry behaves the next time a lab finds itself holding a model it has internally judged to be a step too far.

What is not in dispute is the substance of what Fable 5 and Mythos 5 represent: a meaningful capability jump across coding, reasoning, vision, and scientific research, deployed alongside a safety architecture that is more granular and more publicly documented than anything Anthropic has previously shipped at this scale. The two-model structure, one model with the safety brakes engineered in as a first-class architectural feature rather than an afterthought, and one model with those brakes selectively released to people whose job is to use the capability for defence rather than attack, may turn out to be less a one-time release strategy and more the template for how the next generation of frontier models reaches the public at all.