Anthropic made a decision no AI company wants to make: their flagship model went live, and within 72 hours, access was cut off for everyone.
Claude 5 (internally codenamed Fable 5 / Mythos 5) launched in early June 2026. Users flooded in. Developers started integrating the API. Then it stopped.
What Happened
The US government intervened, citing evidence of jailbreaking methods that could bypass Claude 5's safety mechanisms. Rather than patching specific vulnerabilities, Anthropic chose to take down the model completely — not just for foreign users as the initial order suggested, but for all customers globally.
Active conversations were cut mid-session. New sessions defaulted back to Opus 4.8. No advance notice. No announcement of refund policy.
Two Ways to Read This
Reading 1: Calculated brand management. Anthropic's founding narrative is "safety-first AI." Showing any resistance to government intervention would undermine that story. Short-term revenue loss, long-term trust preserved.
Reading 2: A systemic problem exposed. When jailbreaking techniques spread faster than model iteration, no company can guarantee a "safe launch." Claude 5 was just the first to get caught.
The timing matters: the same week Claude 5 was taken down, Zhipu AI in China announced the full open-source release of GLM-5.2 — moving in the exact opposite direction.
The User Rights Question
Users who paid for Claude Pro subscriptions were buying access to the latest models. When Claude 5 was pulled and replaced with Opus 4.8, did Anthropic breach that contract?
No refund policy was announced. Communications focused on "cooperating with the investigation," not apologizing to customers.
This exposes the blurriest part of AI service agreements: you're not buying a fixed product. You're buying "access to Anthropic's currently available models" — a definition that can be changed unilaterally.
What Comes Next
Claude 5's 72-hour window may be a landmark moment — the first time a major AI model was pulled not for technical failures, but for regulatory intervention.
As models get more capable, the gap between "what they can do" and "what governments will allow" is shrinking. Claude 5 was the first. It won't be the last.
Read the full analysis on my blog: Claude 5下线完整解析
Top comments (0)