Anthropic launches Claude Gov models for national security customers

Anthropic announced Claude Gov, a family of models built specifically for US national-security customers and already deployed, the company said, by agencies operating at the highest levels of classified access. The models were developed from direct feedback from government customers and were, per Anthropic, already in use before the public announcement — an unusual sequencing for a model release, and one that underscored how closely the company had been working with intelligence and defence agencies ahead of the launch.

The stated differences from Anthropic’s consumer and enterprise Claude models centred on refusal behaviour: Claude Gov models decline fewer requests when engaging with classified material, reflecting that legitimate national-security use routinely requires processing content that would trigger safety refusals in a general-purpose deployment. Anthropic said the models also showed improved comprehension of intelligence and defence-specific documents and cybersecurity data, and better handled specialised terminology common in those domains. The company said the models underwent the same safety testing regime as its other Claude releases, an assurance offered because the relaxed refusal behaviour was the release’s most safety-relevant feature.

The launch placed Anthropic alongside OpenAI, which had pursued its own government and defence contracts, in explicitly courting national-security customers with purpose-built model variants — a shift from frontier labs’ earlier public emphasis on broad safety restrictions toward tiered products where restrictions vary by customer clearance and use case. Access to Claude Gov models was limited to government customers operating in classified environments; Anthropic directed interested agencies to its public-sector team rather than offering general availability.