Microsoft, Google and xAI give US early access to frontier AI models

Giving a US federal lab routine, pre‑release access to frontier models is another step toward institutionalizing safety reviews at the same pace that capabilities advance. For Race to AGI readers, this is a concrete sign that model evaluations are moving from voluntary pledges into quasi‑regulatory practice, at least for the biggest US‑aligned labs. It also formalizes CAISI as a gatekeeper for how much governments really understand what’s inside cutting‑edge systems.

Strategically, the deal tightens the coupling between Washington and the current frontier labs: Microsoft as OpenAI’s main platform, Google with Gemini, and xAI with Grok. Smaller players and open‑source projects are conspicuously absent, which could entrench a two‑tier ecosystem where only a handful of firms shape how “secure” AI is defined. That has implications for which architectures, training regimes and red‑teaming practices become de facto standards.

Competitively, this arrangement deepens the moat for the companies already closest to the US government. If CAISI’s tooling, datasets, and test harnesses are tuned around their models, future policy could implicitly favor similar designs. For the race to AGI, it doesn’t directly change the pace of research, but it does signal that frontier work will increasingly run in lockstep with national‑security institutions, not just markets.