Anthropic just made its most powerful artificial intelligence publicly available, but it came with a massive catch.
On June 9, 2026, the company dropped two new models built on the exact same architecture: Claude Fable 5 and Claude Mythos 5. If you've been following the AI industry lately, you already know the backstory. The underlying "Mythos" technology leaked months ago, causing a minor panic in Washington over its frighteningly good offensive cybersecurity and bioweapons capabilities. The White House even called an emergency meeting with financial leaders to discuss the risks.
Instead of keeping the tech locked in a vault forever, Anthropic chose a different path. They took the raw, dangerous model, wrapped it in a heavy outer layer of safety filters, and named it Fable 5 for the general public. Meanwhile, the uncensored version—Mythos 5—is going straight to government-backed defense teams.
It’s a bizarre strategy. It essentially creates a two-tiered system for the absolute peak of machine intelligence.
The Raw Power and the Handoff Mechanism
Fable 5 is an absolute beast of a model. It comfortably beats OpenAI's GPT-5.5 and Google's Gemini 3.1 Pro across software development, scientific data analysis, and long-form logical reasoning.
To give you an idea of what that looks like in the real world, fintech giant Stripe got early access to Fable 5. They used it to completely migrate a massive, 50-million-line Ruby codebase. The model finished the entire project in a single day—a task Stripe estimated would take a human engineering team more than two months.
But you can't use that raw power for everything. Anthropic built Fable 5 with a strict safety handoff mechanism.
When you ask Fable 5 to do something related to hacking, chemical formulas, or extracting its own training data, a hidden classifier triggers. Instead of answering you, Fable 5 instantly routes your request to Claude Opus 4.8, an older, weaker, and safer model.
[User Prompt] ---> [Fable 5 Safety Classifiers]
|
+---> Triggered? YES ---> [Reroute to Claude Opus 4.8]
|
+---> Triggered? NO ---> [Process by Full Fable 5 Engine]
Anthropic admits these safety filters are tuned conservatively right now. They're going to block harmless prompts by accident. In fact, internal data shows the safety handoff will disrupt your normal conversations just under 5% of the time. If you're trying to write legitimate software security code or analyze biological data, expect to hit a wall frequently.
What Happens to the Gated Version
While regular developers play with the filtered Fable 5, the raw, unfiltered tech is moving into a very different circle.
Claude Mythos 5 doesn't have the Opus 4.8 safety handoff. It has full, unmitigated access to its own advanced reasoning capabilities in high-risk areas. Because of that, Anthropic is treating it like a digital hazardous material.
Right now, Mythos 5 is strictly limited to Project Glasswing, a highly vetted partnership that includes the US government, Apple, and NVIDIA. These groups are using the raw model to defend critical infrastructure and find security vulnerabilities before malicious actors do.
Anthropic says Mythos 5 doesn't quite hit the threshold of creating entirely novel biological weapons. But they openly admitted in their official system card that the model can significantly uplift well-resourced threat actors if it ever gets out. It marks a strange era where the public gets the safe family car, while the military gets the tank built on the exact same chassis.
The True Cost of Frontier Intelligence
If you want to use this new tier of intelligence, you better have deep pockets. Fable 5 and Mythos 5 are not cheap.
Anthropic set the price at $10 per million input tokens and $50 per million output tokens. That’s exactly double the cost of running Claude Opus 4.8. For smaller startups and independent developers who are already struggling with the mounting costs of AI infrastructure, this price tag is going to be a massive barrier.
There's a temporary break for casual users, though. Anthropic is offering Fable 5 for free to Pro, Max, Team, and Enterprise subscribers until June 22, 2026. It’s a smart move to let people test the waters and see if the performance boost justifies the premium. But on June 23, the free ride ends. After that, you'll have to pay out of pocket using usage credits until Anthropic scales up enough server capacity to include it in standard subscription plans again.
Practical Next Steps for Builders
If you're managing software teams or building AI-native tools, don't just stand on the sidelines. Take advantage of the free testing window before June 22 to audit your current pipelines.
- Audit your complex tasks: Run your most resource-heavy, multi-step engineering prompts through Fable 5 right now to see if the time saved offsets the 2x pricing increase.
- Test the guardrails: If your application relies on medical documentation, security auditing, or code generation, stress-test your existing prompts. You need to see if they accidentally trigger the 5% fallback classifier to Opus 4.8, which could break your app's user experience.
- Implement billing fallbacks: If you update your API to
claude-fable-5, update your error-handling code to handle potential refusal responses gracefully so your system doesn't freeze when the safety filters trigger.
