OpenAI and Anthropic will give the U.S. government early access to their frontier models for safety evaluation.