Since Anthropic's public rollout of Claude Fable 5 on June 9, the coverage has focused heavily on benchmark performance and IPO positioning. The overlooked story involves the deliberate removal of Mythos's most powerful — and most dangerous — capability before general release.

What Got Stripped Out

The Mythos preview, earlier this year, found thousands of software vulnerabilities on its own. That was a demonstration of AI-enabled offensive cybersecurity at a scale no prior public model had shown.

The public Claude Fable 5 will NOT do that.

Anthropic's head of product management, research, and labs, Dianne Penn, explained the restriction to Reuters with a practical example: a college student asking Fable 5 to find vulnerabilities in a specific code package gets refused. The model then falls back to Opus 4.8 for a safer response. The full Mythos cybersecurity capability stays locked behind the Glasswing program — roughly 200 vetted organizations, including the U.S. government.

Anthropic is making a hard call that the dual-use risk of public cyberweapon-capable AI outweighs the commercial upside of releasing it broadly. It's a consequential safety decision.

The Commercial Logic Alongside the Safety Logic

This also serves Anthropic's business interests. Anthropic is a $965 billion company, according to Reuters, racing OpenAI to a public offering. Keeping the most powerful Mythos capabilities gated behind a "systematic trusted-access program" maintains premium pricing power and reduces regulatory risk heading into an IPO roadshow.

Pricing on both public Fable 5 and the restricted Mythos 5 is set at $10 per million input tokens and $50 per million output tokens, according to Anthropic. Penn told Reuters that despite Fable 5 being a more expensive model, it accomplishes tasks with lower token usage — meaning the effective cost per task is actually lower, based on early customer feedback.

This pitch directly addresses the model-routing trend identified in June 5 coverage, where corporate America is increasingly choosing cheaper, task-specific models over expensive flagship ones.

The Strongest Counterargument: Is Blocking Cybersecurity Capabilities Enough?

Security researchers argue that the real threat isn't a college student asking Claude for help — it's sophisticated state actors and criminal organizations who will probe, jailbreak, and systematically test any public model for exploitable gaps. Anthropic says it has done "extensive testing to ensure that users could not manipulate the new model to bypass its guidelines," but that claim warrants scrutiny.

Every major AI guardrail to date has eventually been partially circumvented. Anthropic's assurance that Fable 5 won't be manipulated into performing restricted actions is significant. The company has NOT published independent third-party red-team results to back it up publicly, and no external security audit has been cited in these sources.

Anthropic's defenders counter that a soft refusal falling back to a safer model is architecturally sounder than a hard block that might create adversarial pressure to work around the restriction. If the model gracefully degrades rather than hard-refusing, it's harder to probe for a bypass. That's a defensible design philosophy — but it's also unproven at scale.

The Governance Question

Most outlets are running this as a straightforward product launch story. The more critical story is the governance architecture Anthropic is quietly building.

The Glasswing program — 200 organizations with privileged access to uncapped Mythos — is essentially a private regulatory tier operating outside any formal government framework. The U.S. government is a participant, not an overseer.

Who decides which organizations get Glasswing access? What are the vetting criteria? What oversight exists if a Glasswing participant misuses the full Mythos capability? Anthropic's public announcements don't answer these questions, and few reporters have pushed for that information.

With a valuation north of $965 billion and a model capable of finding software vulnerabilities at scale, these are governance questions with real consequences. They touch on who controls one of the most capable offensive cybersecurity tools ever built.

The IPO Shadow Over Everything

Anthropic's public Fable 5 rollout, the Glasswing expansion, the $35 billion private credit deal closed by Apollo and Blackstone, Microsoft AI chief Mustafa Suleyman's public warnings about Anthropic's consciousness speculation — all of this is happening in the context of a company preparing for a major public offering.

Every decision Anthropic makes right now carries an IPO dimension. That doesn't make the safety decisions wrong. But investors, regulators, and the public should read every Anthropic announcement with that context clearly in mind.

The public got a powerful AI model with the most dangerous parts disabled. The vetted few keep the full capability. Whether that represents responsible stewardship or a two-tiered power structure that nobody voted for remains the question mainstream coverage keeps avoiding.