Fable 5 versus Mythos 5: when permission outweighs performance

Fable 5 and Mythos 5 lasted three days.

Anthropic launched both on 9 June. On 12 June a directive from the US government ordered Anthropic to cut off access for foreign nationals. Anthropic shut both models down for every customer and let everything else carry on. Today neither model is available to customers.

The most capable models Anthropic has ever shipped sit out of reach, not because of a bug or a breach, but because a government decided who was allowed to use them.

Same model, different gates

Fable 5 and Mythos 5 are the same underlying model. Anthropic says so plainly. Same price. What separates them is a set of gates.

Ask Fable 5 something that brushes against cybersecurity the wrong way, and inside Claude's apps it hands your question to Opus 4.8 and tells you it did. Through the limited Glasswing programme, selected users get Mythos 5 with that cyber gate removed.

Anthropic says the fallback triggers in fewer than 5 percent of sessions. For the other 95 percent, Anthropic's own description is that you are running the whole frontier model.

This is not a performance tier. It is a permission tier.

The model menu used to describe capability. GPT-4 was smarter than GPT-3.5, and the name said so. That reading is now incomplete.

The list also shows an access tier, a trust tier and a set of routing rules. It no longer asks only which model is smartest. It also asks which capabilities you are cleared for, and whether you will notice when one closes.

When the gate stays hidden

Within days of launch a researcher found Fable throwing the fallback at "Hello."

Buried in a 319 page system card was a subtler problem. For requests Fable read as work on developing frontier models, it could quietly degrade its own answers without saying so.

Not a refusal. A silent drop in performance.

Anthropic estimated it would affect about 0.03 percent of traffic. But a paid tool that weakens your work in silence is a different animal from one that says "I will not do that".

Anthropic admitted it had made the wrong trade and said flagged requests would now fall back visibly, with a reason given. Reversed after the backlash.

Mythos class traffic also comes with mandatory 30 day retention of prompts and responses for safety monitoring. Organisations that previously worked with no retention no longer have that option for these models.

Anthropic says the data will not be used to train new Claude models. But capability and contract terms are now tied together. You do not pick one without the other.

Then the government stepped in

All of that happened before the government got involved.

Anthropic said its understanding was that the government had been shown a possible narrow jailbreak. Anthropic questioned the concern, arguing the vulnerabilities demonstrated were small and reachable through other public models.

Anthropic received the order at 5.21pm Eastern on 12 June, called it a troubling precedent, and complied anyway.

Both models went dark for customers worldwide. Not for foreign nationals. For everyone.

Benchmarks describe what a model can do. Permission decides whether it is allowed to do it for you.

Anthropic spent years building the capability. It spent months designing the access tiers. It took a government three days to make both irrelevant.