Anthropic just shipped its most capable AI to the public by making it worse on purpose.
The Summary
- Anthropic released Claude Fable 5, the first Mythos-class model available to the public, but it refuses to answer basic questions about cybersecurity, biology, and chemistry, instead routing those queries to the older Claude Opus 4.8.
- The model isn't dumb, it's deliberately hobbled. Anthropic says Mythos is so good at finding and exploiting software vulnerabilities that releasing it without safeguards could enable serious damage.
- More than 95% of Fable sessions run without triggering the safety fallback, but users asking about cancer spread or basic security concepts are getting kicked to the slower model.
- A separate version, Claude Mythos 5 with safeguards lifted in specific areas, goes to approved cybersecurity partners and biology researchers through Project Glasswing.
The Signal
This is the first time a major AI lab has shipped a broadly capable model to the public while simultaneously admitting they can't trust you with all of it. Fable 5 represents Anthropic's strongest performance across software engineering, knowledge work, vision, and long-running tasks, with the company noting that its lead over other models grows as tasks become longer and more complex. But the model's actual capability set is artificially constrained for everyone except a curated list of approved organizations.
When Business Insider tested Fable 5 with basic questions about cancer, like how misinformation about the disease spreads online or explanations of different cancer types, Claude immediately switched to Opus 4.8 and displayed a notification explaining the safety measures. The pop-up acknowledged what users quickly discovered: "They may flag safe, normal content as well."
"Fable 5 has safety measures that flag messages on most cybersecurity or biology topics. They may flag safe, normal content as well."
The core tension here is between capability and access. Anthropic first announced Mythos in April through Project Glasswing, calling it a "step change" in capabilities but choosing to restrict access to a small group of tech, cybersecurity, and financial companies plus the US government. Two months later, they've found a way to make it broadly available by creating two versions of the same underlying model with different names and access policies.
The approved users getting the unrestricted Claude Mythos 5 include cyberdefenders, infrastructure providers, and select biology researchers. Everyone else gets Fable 5, which Anthropic describes as having "the same underlying Mythos-class capability" wrapped in safeguards that automatically route high-risk queries to the previous generation model. The company says requests involving cybersecurity, biology and chemistry, and model distillation trigger the fallback.
Key safeguard details:
- Over 1,000 hours of internal and external red-teaming found no universal jailbreaks
- 95%+ of sessions run entirely on Fable without triggering fallback
- Users get notified when the model switches to Opus 4.8
Anthropic's stated justification is that without safeguards, Fable 5's capabilities in cybersecurity "could be misused to cause serious damage." The implication is that Mythos can find vulnerabilities in critical software that current models miss. That's a capability valuable enough that Anthropic wanted to ship it to approved partners months ago, but dangerous enough that the public version needed an automatic downgrade switch for entire categories of queries.
The Implication
We're entering an era where the best models won't be available to everyone with the same capabilities. Anthropic chose broad access with constraints over restricted access with full capability. That's probably the right call for now, but it creates a two-tier system where approved organizations get tools the rest of us can only read about.
For developers building on Claude, this matters immediately. If your application involves cybersecurity workflows, biological research, or chemistry, you're not actually building on Mythos-class capability unless you get approved for Glasswing access. You're building on Opus 4.8 with extra latency. The model you're testing against today might not be the model your users get tomorrow if your prompts trip the classifier. Anthropic says they're working to refine the safeguards, which means the boundary of what triggers fallback will shift over time.
Watch for competitors to face the same choice. If OpenAI or Google releases something comparable, they'll need to decide whether to gate access entirely or ship a hobbled public version. The Fable approach lets Anthropic claim they've democratized Mythos while maintaining control over the capabilities that actually make it dangerous. That's clever product strategy, but it's not the same as open access.
Sources
The Verge AI | Business Insider Tech | Latent Space | Mashable Tech | VentureBeat | Bloomberg Tech | TechCrunch AI | Fortune Tech