US government pre-deployment AI review expands from two labs to all five frontier developers
Before
As of August 29, 2024, only Anthropic and OpenAI had Memoranda of Understanding with the US AI Safety Institute (AISI) granting the government access to major new models before and following their public release. Google DeepMind, Microsoft, and xAI operated without US government pre-deployment evaluation agreements.
→ HPCwire — US AI Safety Institute Collaborates with Anthropic and OpenAI on AI Safety Research (NIST press release)After
On May 5, 2026, the Center for AI Standards and Innovation (CAISI) — the Trump administration's successor to AISI under the Department of Commerce — announced pre-deployment evaluation agreements with Google DeepMind, Microsoft, and xAI. All five major US frontier AI developers now submit unreleased models to government evaluation before public launch. CAISI has completed more than 40 such evaluations, including on models that remain unreleased.
→ HPCwire — NIST's CAISI Announces New Frontier AI Testing Agreements with Google DeepMind, Microsoft, xAI (NIST press release)Consequence
Every major model release from a US frontier AI lab — GPT-*, Claude, Gemini, Grok — now goes through a government evaluation window before it reaches developers and enterprises. The practical effect for builders is slightly longer gaps between model announcement and API availability. The structural effect is that the US government has visibility into unreleased frontier model capabilities, including evaluations conducted in classified environments with reduced or removed safeguards, before those capabilities are public.
The expansion is notable given the administration's context. The agreements were renegotiated under CAISI's mandate from the Secretary of Commerce and the America's AI Action Plan — a Trump-era document from a White House that has publicly prioritised AI competitiveness over AI safety regulation. The program started under Biden-era commitments with two labs; it now covers all five under a different administration. The agreements remain voluntary — no lab is legally obligated — but all five have signed.