The Lion at the Door
The Lion at the Door
There is a useful distinction that tends to get lost in the coverage of AI model releases. Most of what gets discussed — Claude, ChatGPT, Gemini — belongs to a category of tool that is, at its core, a general-purpose assistant. It answers questions, writes drafts, explains concepts, helps with code. It lives comfortably in the hands of anyone with a browser and a few minutes. It is, to borrow an image, the rabbit that lives in your house. Friendly. Useful. Occasionally surprising. Fundamentally domestic.
What Anthropic released today is not that.
Claude Fable 5, launched Tuesday morning, is the first publicly available version of the Mythos model family — a class of AI that the company itself kept locked behind a restricted program for months because of what it could do in the wrong hands. This is the animal outside the house. It is not bigger than the rabbit. It is a different kind of creature entirely.
What Mythos Actually Is
To understand why Fable 5 matters, you have to understand what Mythos was built to do. Most AI models are trained to be broadly capable — good at a wide range of tasks, constrained from the most dangerous ones. Mythos was built with a specific capability at its core: finding the holes in software that nobody has found yet.
In the field of cybersecurity, these are called zero-day vulnerabilities — flaws in operating systems, browsers, and critical infrastructure that have not yet been discovered or patched. Finding them is extraordinarily difficult, time-consuming work. It requires the ability to hold an entire complex software system in working memory, reason about how its components interact, identify edge cases that human engineers missed, and chain multiple small weaknesses into a path that an attacker could exploit.
Mythos can do this autonomously. During Project Glasswing — the restricted program Anthropic launched in April 2026 with partners including AWS, Microsoft, Apple, and CrowdStrike — the model identified more than ten thousand high- and critical-severity vulnerabilities across globally significant software systems in its first weeks of operation. The bottleneck in cybersecurity defense, the When AI Builds Itself paper noted, had already shifted from finding vulnerabilities to patching them fast enough.
That is weapons-grade capability. Calling it anything softer — security-grade, infrastructure-class, advanced cybersecurity tooling — would be like looking at a lion and calling it a large house cat. Technically in the same family. Categorically not the same thing.
Calling it anything softer would be like looking at a lion and calling it a large house cat. Technically in the same family. Categorically not the same thing.
What Fable 5 Actually Is
Fable 5 is Mythos with the cage door partially closed.
Technically, Fable 5 and Mythos 5 run on the same foundation — the same underlying model, the same training, the same core capabilities. What separates them is a layer of classifiers that Anthropic has built on top of Fable 5. When a query touches high-risk domains — cybersecurity exploitation, biology, chemistry — the classifiers intercept it and route it to Claude Opus 4.8 instead. The weapons-grade capability is still in there. The safeguards are designed to prevent it from being accessed directly.
Mythos 5, released simultaneously, is the same model with fewer of those restrictions. It goes only to a small group of vetted cyberdefenders and infrastructure providers — organizations that already had access to the original Mythos Preview through Project Glasswing.
The distinction matters. Fable 5 is not a neutered version of Mythos. It is Mythos with specific, deliberate guardrails applied to the most dangerous outputs. Whether those guardrails hold under adversarial pressure — whether determined users can find ways around the classifiers — is a question the field will be watching closely.
The Price of Entry
Weapons-grade tools carry weapons-grade price tags. Fable 5 is priced at $10 per million input tokens and $50 per million output tokens — exactly double the cost of Claude Opus 4.8, and still the most expensive major AI model available to the general public globally. The original Mythos Preview, available only to Glasswing partners, was priced at five times the Opus rate. Today's pricing represents a significant reduction from that, but it still positions these tools well above the consumer market.
The pricing is not incidental. It is a screening mechanism. At double the cost of the current flagship model, Fable 5 is not something a casual user picks up on a whim. It targets institutions, enterprise security teams, and developers building serious infrastructure — the population most likely to have a legitimate use case and the accountability structures to go with it.
For subscription users, Anthropic is offering Fable 5 at no extra cost through June 22 on Pro, Max, Team, and Enterprise plans — a limited window before usage credits become required. The company says it aims to restore Fable 5 as a standard subscription feature as quickly as possible, which suggests the current access model is a temporary arrangement while capacity scales.
The Parallel Track
Anthropic is not alone in this territory. In May 2026, OpenAI released GPT-5.5-Cyber — a variation of its latest model, rolled out in limited preview to vetted cybersecurity teams. The two approaches tell different stories about how to handle the same problem.
OpenAI's philosophy, as it has described it publicly, is democratization with verification: make these tools as widely available as possible, control access through identity verification and monitoring rather than capability restriction. The model itself is not constrained — the question is who gets to use it and under what conditions.
Anthropic's approach, at least until today, has been the opposite: capability restriction first, controlled expansion second. Mythos stayed behind closed doors for months. Fable 5 is the first step outward, and it comes with classifiers built into the model itself rather than gates around it.
Neither approach has been tested at scale. Both are reasonable responses to a genuinely hard problem. The hard problem is this: the same capability that lets an AI system find a zero-day vulnerability in critical infrastructure for a defender can find the same vulnerability for an attacker. Defensive and offensive cybersecurity use the same knowledge. There is no version of a weapons-grade AI that is useful only to the good guys. The capability does not know whose hands it is in.
There is no version of a weapons-grade AI that is useful only to the good guys. The capability does not know whose hands it is in.
What Fable 5 Can Do Beyond Cybersecurity
The cybersecurity capability is what earned Mythos its classification. But Fable 5's capabilities extend well beyond it. Anthropic describes the model as exceptional across software engineering, knowledge work, vision, scientific research, and long-running autonomous tasks. The longer and more complex the task, the larger Fable 5's lead over previous models.
In life sciences research, Mythos-class models have been positing novel hypotheses and accelerating the development of new therapeutics — a domain where the bottleneck has historically been the speed at which researchers can generate and test ideas. In software engineering, Fable 5 scored 80.3 percent on SWE-Bench Pro, the standard real-world benchmark for autonomous code repair, compared to 69.2 percent for Opus 4.8 and 58.6 percent for GPT-5.5.
The model can also work autonomously for longer than any previous Claude — handling tasks that span days rather than hours, maintaining context and making decisions across extended timelines without human check-ins at each step. That autonomous duration is itself significant. The longer a model can operate independently on a complex task, the more it begins to resemble something closer to a colleague than a tool.
The Context Nobody Should Miss
Fable 5 launched five days after Anthropic published When AI Builds Itself — a document that called for international coordination to slow or pause frontier AI development before recursive self-improvement becomes possible. It launched one week after Anthropic filed confidentially for a public offering that values the company near a trillion dollars. And it launched as the most powerful publicly available AI model in the world, in a category that did not exist eight months ago.
None of those facts contradict each other. A company can genuinely believe that AI development is moving faster than governance can keep up, advocate for coordination mechanisms, and still release its most capable model to the public — because the alternative is ceding that capability entirely to actors with fewer scruples about the questions it raises. That is a real argument. It is also the argument every actor in every arms race has ever made.
The technology is here. The question of what to do with it is not settled. That question belongs to more than the companies building these systems — and that, more than any benchmark score or pricing tier, is what today's release actually puts on the table.
The full arc — from the 2017 Transformer paper to recursive self-improvement and the three futures Anthropic says are coming. Read "The Idea That Built the Age" at Tech Reader Magazine.