One model, two very different sets of rules
Anthropic's newest release reads less like a story about raw capability and more like one about restraint. The company has shipped two flavors of the same underlying system: Claude Mythos 5, reserved for a narrow circle of trusted partners, and Claude Fable 5, opened up to the general public. That division gets at a problem Anthropic still hasn't fully cracked — how do you put an extraordinarily capable model into the world without handing would-be attackers a fresh set of offensive tools at the same time?
The short answer is that you don't release the unleashed version to everyone. You build a constrained sibling, and you let that one go public.
What the Mythos preview already proved
Mythos has shown what it's capable of when the restraints come off. Beginning in April, an earlier preview went out to roughly 150 organizations under a program Anthropic calls Project Glasswing. In the months since, those users have flagged more than 10,000 critical security vulnerabilities in their own systems — the kind of find that lets defenders patch holes before anyone else walks through them.
But that's the catch. The very same skill that uncovers a flaw so it can be fixed is the skill an attacker would use to exploit it. Capability doesn't pick sides.
Why Mythos 5 is staying behind the glass
Because of that dual-use reality, Mythos 5 isn't going wide. Anthropic is keeping it in the hands of a small group of cyberdefenders and infrastructure providers, alongside a selection of biology researchers, and it's coordinating with US government agencies as the rollout proceeds. Access runs on a need-to-know basis. The company has signaled that a broader "trusted access program" will follow at some later point, but for now the door stays mostly shut.
How Fable 5 is fenced in for public use
Fable 5 is Anthropic's experiment in what a general-release version of Mythos-class technology can look like once you wrap it in constraints. Underneath, it's the same model as Mythos 5. The difference is the hard limits bolted on top.
Guardrails that quietly reroute risky prompts
Fable 5 is built to refuse or redirect a long list of requests touching cybersecurity, biology, and chemistry. Here's the part that's easy to miss: when one of those guardrails fires, the query doesn't simply get rejected. It's silently handed off to an older model, Claude Opus 4.8, which answers in Fable 5's place. More often than not, the user won't even realize the swap happened.
Protection against model distillation
Anthropic also wired Fable 5 to watch for distillation — the practice of vacuuming up huge volumes of answers to train a smaller, cheaper model that mimics the original. If the system suspects someone is doing exactly that, it reroutes those requests to Opus 4.8 as well. So Anthropic isn't only policing what the model will discuss; it's guarding what anyone can extract and learn from it.
Tuned to over-block, on purpose
Right now, the filters lean hard toward caution. They're set to over-block, which means a fair number of perfectly harmless questions get bumped to the older model anyway. Diane Penn, who leads product management at the company, has been candid that the setup is far from perfect, and that feedback gathered since the April preview is what shaped the current approach.
Her framing is pragmatic. The goal, she's said, is to make genuinely useful improvements even without a flawless answer for every scenario. Out of the options on the table, this one looked like the most workable — and the choice most likely to let people pull real value out of Fable 5. Anthropic plans to sharpen its classifiers over time, but it argues that this degree of caution is the price of justifying a wider release at all.
Agentic behavior raises the stakes
What makes this more than an academic debate is that Fable and Mythos aren't simple chatbots that answer a prompt and stop. Anthropic says both can run unattended for longer stretches than earlier Claude models, working through sequences of instructions without someone hovering over every step.
That shift toward more agent-like behavior could meaningfully speed up software engineering and other technical work, especially with Fable 5's sharper code generation and stronger visual capabilities. The flip side writes itself: the more a system can do on its own, the more there is to worry about if someone points it in the wrong direction.
What Fable 5 and Mythos 5 cost
The pricing tells you how much horsepower Anthropic believes it's selling. Both Fable 5 and Mythos 5 run $10 per million input tokens and $50 per million output tokens — roughly twice what the company charges for its other public models, though still a step down from the earlier Mythos Preview. That premium reflects both the performance jump and the fact that these are still being treated as specialized systems rather than just another entry in an ever-growing catalog.
How rivals and regulators are reacting
Anthropic isn't moving in isolation. OpenAI has pushed out its own advanced cybersecurity model to a small set of partners and stood up a working group that closely mirrors Project Glasswing. Both companies are eyeing potential public offerings, which adds its own pressure — they need to convince investors they can ship frontier technology without setting off a backlash over safety.
And the unease isn't confined to industry insiders. Canada's finance minister, François-Philippe Champagne, has said publicly that the worry surrounding Mythos really comes down to fear of what nobody can see coming yet — the unknowns you don't even know to look for.
Anthropic co-founder Jack Clark has made a strikingly similar point from inside the company, arguing that the industry still hasn't worked out how to slow itself down when it needs to. His metaphor lands hard: the AI field has built itself a gas pedal but never bothered to install a brake — and you really want the option to ease off the accelerator and stop when the moment calls for it.