Every major AI lab in the country has quietly given Washington a preview of its most powerful models before they ship. OpenAI, Anthropic, Google, Microsoft, xAI — all in. Then there’s Meta, still outside the velvet rope, negotiating terms while the clock runs down. A June 2 executive order established a voluntary framework for pre-release government review of “covered frontier models,” granting federal agencies up to 30 days to evaluate systems before they reach trusted partners or the public. The formal process won’t be finalized until late July. But “voluntary” is doing heavy lifting here — because the government just showed everyone what happens when it stops asking nicely.
The Voluntary Framework That Isn’t Really Voluntary
The policy landscape has already moved past debate — most major AI firms are sharing models before any formal requirement exists.
Here’s where things stand right now, according to reporting from the New York Times and Reuters:
- Google, Microsoft, and xAI have agreed to give the Commerce Department’s new Center for AI Standards and Innovation early access to unreleased models for safety testing and capability assessments.
- OpenAI and Anthropic are already working directly with federal agencies on pre-release evaluations.
- Meta is the only major U.S. AI developer that hasn’t signed on. Government outreach reportedly arrived via email.
- Most firms began sharing models informally before the executive order even existed.
“We share the administration’s goal of advancing U.S. leadership on robust and secure frontier AI. While we are working through the details, we hope to sign the agreement soon.” — Meta spokesperson, via Reuters.
Meta isn’t refusing. It’s negotiating scope, intellectual property protections, and how its open-source Llama models fit a framework designed for closed systems. That distinction matters more than the headline suggests.
Muse Spark and the Model That Changed the Conversation
Meta’s newest closed-weight model puts it directly in the same regulatory crosshairs as its rivals — regardless of its open-source track record.
Meta launched Muse Spark in April 2026 as its most capable model yet, embedded across its consumer apps with two reasoning modes: “Instant” for speed, “Thinking” for deeper problem-solving. Independent benchmarks place it below leading GPT and Gemini variants on coding and agentic tasks, but notably strong in health and medical reasoning. Unlike Llama, Muse Spark is closed-weight — which lands it squarely in regulators’ sights. Readers curious about how AI-Powered Websites are reshaping everyday productivity can find that these same consumer-facing integrations are now drawing regulatory scrutiny.
The sharpest signal came in mid-June. The administration ordered Anthropic to suspend all foreign national access to its Mythos 5 and Fable 5 models over cybersecurity concerns, according to the New York Times. Anthropic responded by killing access for everyone to ensure compliance. That’s the real message — a framework labeled “voluntary” escalated into a complete shutdown in a matter of weeks.
Meta’s open-source credibility helped Llama earn government-wide approval through the GSA’s OneGov initiative. That goodwill doesn’t extend to Muse Spark. If Meta signs on, the U.S. effectively has a de facto national review regime covering every major frontier model — and the gap between voluntary oversight and binding regulation has now been measured in weeks, not years.
