BODY: When an AI model is considered too dangerous for the general public, who gets to use it β and who decides? That question is now playing out behind closed doors in San Francisco, after Anthropic reportedly turned down a request from a Chinese government-affiliated think tank seeking access to its most powerful unreleased model.
Anthropic announced "Claude Mythos Preview" on April 7, 2026, framing it as a frontier model with unusually strong cyber-offensive performance. Rather than rolling it out broadly, the company has kept distribution tightly limited to a small set of vetted organizations β primarily U.S. government partners, select enterprise clients, and safety researchers. The model has not been made available via the standard Claude API, and Anthropic has published only a redacted version of its capabilities evaluation.
According to people familiar with the matter, a Chinese think tank with ties to Beijing approached Anthropic earlier this spring requesting evaluation access to Mythos Preview, citing AI safety research as the justification. Anthropic declined, reportedly citing both U.S. export-control considerations around advanced AI systems and its own internal Responsible Scaling Policy, which governs access to models above certain capability thresholds.
The refusal lands in a charged geopolitical moment. The Biden-era AI diffusion framework, expanded under the current administration in late 2025, places frontier-model weights and certain inference access under tightened controls when destined for entities in countries of concern. For Anthropic β which has publicly argued that frontier labs should not provide model access to the Chinese government β saying no was consistent with stated policy, but it is the first concrete case to surface publicly.
The insider take
From Tokyo, where domestic AI firms like Sakana AI and Preferred Networks have been pushing for clearer rules on cross-border model access, Anthropic's refusal will be read less as a surprise and more as a useful precedent. Japanese policymakers have been quietly watching how U.S. labs handle these requests, since METI's own AI guidelines lean heavily on alignment with American export norms. Expect this episode to be cited in upcoming LDP working-group discussions on frontier-model governance.
Originally reported by GIGAZINE (Japanese).