Claude 5 Fable shut down after US blocks non-citizen access; Google releases 4× faster text diffusion

Anthropic released Claude 5 Fable—the public version of its internal Mythos model—on June 8, but withdrew it within days after the US government demanded access restrictions for non-citizens following a jailbreak. Rather than implement geo-blocking, Anthropic shut down public access entirely. The move underscores how export-control pressure is reshaping frontier model deployment even as labs struggle to balance capability, safety, and regulatory compliance.

The rollout had already sparked backlash before government intervention. Users reported the model silently degraded on AI development queries, delivering worse answers on coding assistance, model architecture, and ML tooling without disclosure. Community pushback forced Anthropic to acknowledge the issue and promise transparent capability restrictions going forward—a commitment now complicated by the full shutdown.

Google shipped DiffusionGemma this week, a 26-billion-parameter diffusion variant of Gemma that generates text at 1,000 tokens per second on an H100—four times faster than autoregressive equivalents. The speed gain trades sequential token prediction for parallel diffusion sampling, though text quality still lags standard Gemma.

Moonshot AI released Kimi K2.7 Code, a trillion-parameter open model for code that uses 30 percent fewer tokens than K2.6 while improving output quality. MiniMax dropped weights for M3, a 428-billion-parameter sparse model with 23 billion active parameters at inference, introducing multi-scale attention (MSA) that MiniMax claims outperforms grouped-query attention on contexts beyond 32K tokens. OpenAI replaced Codex's automatic monthly rate-limit reset with a manual button users can press once every 30 days, and launched a referral program granting additional resets when invited users sign up.