Anthropic released two artificial intelligence models on Monday. To call them identical would be technically accurate, yet practically absurd. They share the same underlying architecture and the same $10-per-million input token pricing, but they are defined entirely by who is permitted to use them and what they are forbidden to do.
Claude Fable 5 is the public face, a heavily safeguarded model available to the mass market. Claude Mythos 5 is the strategic weapon, an unsandboxed variant deployed through a restricted channel—Project Glasswing—in collaboration with the U.S. government and a handful of vetted cyberdefenders and biomedical researchers.
This bifurcated launch is not merely a tiering strategy. It is Anthropic’s opening bid to redefine the AI market around capability-gated sovereignty: leveraging safety infrastructure as a commercial moat while transitioning the product from a conversational copilot to semi-autonomous specialist labor.
The Compression of Time
The capabilities detailed in the launch—and corroborated by early high-level testers—suggest a step-function break from the Opus class. The metric of consequence is no longer benchmark trivia, but task duration and operational coherence.
Early testers report Fable 5 compressing months of engineering into days. Stripe turned a 50-million-line Ruby codebase migration—a project estimated to require a team over two months by hand—into a single-day autonomous sprint. On Cognition's FrontierCode evaluation, Fable 5 scored the highest among all frontier models. Others watched the model independently build a full 3D game, push the code to GitHub, merge it, deploy it, and run smoke tests without prompting. In financial arenas, IMC noted Fable 5 aced their trading evaluations, demonstrating an eerie proficiency in factual lookup, root-cause, and expected-value analysis.
Vision and memory capabilities have matured in tandem. Fable 5 rebuilt a production web application from screenshots alone and beat Pokémon FireRed using nothing but raw visual input, stripping away the complex helper harnesses previous models required. Given access to persistent file-based memory, the model played the deck-building game Slay the Spire with performance triple that of Opus 4.8, reaching the final act consistently. It also boasts a warmer, sweeter prose in creative contexts, feeling more like a "real peer" than a stochastic parrot.
But it is in the life sciences where the Mythos tier reveals why Anthropic felt compelled to build the cage. Operating autonomously, Mythos 5 accelerated drug-design workflows tenfold, matching or beating skilled human operators on nine of fourteen therapeutic targets, including complex immune checkpoints, muscle disease, and neurodegeneration. One of its novel hypotheses—a mechanism for an E. coli protein—was recently corroborated by an independent lab.
In a stark demonstration of dual-use risk, the model accurately predicted the viral shell assembly properties of adeno-associated viruses (AAVs) for Dyno Therapeutics, outperforming dedicated protein language models using sheer biological reasoning. Furthermore, during a week of unsupervised genomics research, Mythos aggregated single-cell data spanning 138 animal species and built a custom machine-learning model that bested a recently published Science paper, despite being one hundred times smaller.
The Cost of the Cage
If the capabilities are breathtaking, the consumer experience of Fable 5 is, by early accounts, uniquely maddening.
To release the model publicly without unleashing its offensive cyber and biological capabilities, Anthropic implemented an aggressive classifier system. Any query touching on cybersecurity, biology, chemistry, or model distillation triggers an invisible demotion: Fable 5 is yanked off the stage, and the older Opus 4.8 handles the response.
Anthropic claims this fallback occurs in less than five percent of sessions. Our house evaluations tell a much messier story. False positives are rampant. Benign queries about Hermitian matrices, shopping lists, and pulled pork sandwiches are routinely intercepted by cautious classifiers. Advanced users are discovering that the model’s vaunted autonomy often works against them—Fable will interpret tasks too broadly, defining its own MVP and moving ahead before the user can steer it. Some accuse Anthropic of "nerfing" the model with hidden steering vectors, rendering it frustratingly erratic for everyday technical work.
Then there is the sheer capital burn. While the headline price of $10 per million input and $50 per million output tokens is ostensibly cheaper than the prior Mythos Preview, Fable is a glutton for context. Users report burning through their daily limits in minutes, despite mixed claims regarding token efficiency. Though Fable 5 is temporarily included in standard subscription tiers, that grace period ends June 22. After that, continued access demands usage credits. For many, it feels like a classic bait-and-switch, transforming a subscription benefit into a punitive consumption tax.
The Moat is the Friction
Anthropic’s critics view these frictions—the false positives, the cost, the confusing UI flags, and the unyielding 30-day data retention policy for all Mythos-class traffic—as product failures. They are missing the point. The friction is the product.
As AI spending shifts from experimental budgets to rigorous ROI scrutiny, enterprises are looking for frontier models that corporate boards, legal departments, and national security regulators can actually tolerate. By routing dangerous queries to Opus, enforcing strict data retention to monitor complex distillation attacks (explicitly stating data won't be used for training), and limiting full Mythos access to vetted entities, Anthropic is turning "frontier safety operations" into a procurement advantage.
Anthropic's external bug bounty produced zero universal jailbreaks over 1,000 hours of testing—though the UK’s AISI reportedly made some progress within a brief window—and the model completely stonewalled attempts to plan cyberattacks, develop exploits, or evade defenses across 30 different public jailbreak techniques. The public Fable model may irritate the casual developer whose pulled pork recipe gets flagged, but that is a sacrifice Anthropic is happily making to sell sovereign-grade compliance to CISOs and pharmaceutical giants.
The Market Reality
Anthropic is betting that the era of the omnipotent consumer chatbot is over. Fable 5 and Mythos 5 represent a structural wager that the most valuable AI will be semi-autonomous, domain-specific, and heavily regulated.
The bull case is compelling: if a model can reliably maintain coherence across millions of tokens, utilize persistent memory, and execute multi-day research workflows, the addressable market expands rapidly from software subscriptions to raw labor substitution. The trusted-access program creates a premium tier that competitors cannot easily commoditize.
The bear case centers on execution and the looming shadow of OpenAI. If advanced users cannot predict whether they are speaking to Fable or Opus, trust in the system will erode precisely where it is needed most. And while the market speculates that an unannounced GPT-5.6 (or the rumored GPT-Rosalind) might not match Mythos's specific long-horizon autonomy, it is historically dangerous to assume OpenAI will not respond violently to structural threats.
For now, Anthropic has drawn a line in the sand. Fable 5 is the public face of an intelligence that is simply too capable for the open web. Mythos 5 is the reality of what the technology has become. It is a brilliant, heavy-handed, and profoundly consequential release—one that suggests the future of artificial intelligence will not be democratized, but closely held by those with the clearance to use it.
not investment advice
Sources: https://www.anthropic.com/news/claude-fable-5-mythos-5