OpenAI Called It a Model Launch. For 12 Days, Washington Called It a Customer List.

Between June 26 and July 9, the most capable model built by the world's most valuable AI lab was available to exactly twenty companies — chosen not by OpenAI, but by the U.S. Department of Commerce.

GPT-5.6 Sol, Terra, and Luna became fully public today. Sol is the flagship, priced at $5 per million input tokens and $30 per million output. Terra matches GPT-5.5's performance at roughly half the cost. Luna undercuts both at $1 and $6. The pricing tells a real story about margin compression at the model layer. It isn't the most interesting thing that happened this week.

The interesting part is the 12 days that preceded launch. Under a June 2 executive order, the Trump administration asked frontier labs to voluntarily hand over up to 30 days of advance access before a public release, so agencies could test for national-security-grade cyber risk. OpenAI complied on a compressed timeline: the Commerce Department's Center for AI Standards and Innovation ran its own evaluation, and Commerce Secretary Howard Lutnick personally signed off on roughly 20 named customers who could use GPT-5.6 before anyone else could.

Axios reported this as the administration giving OpenAI a "green light." The White House pushed back within hours, telling reporters it had given no approval, clearance, or permission — that the decision to release "rests entirely with the companies," and that the executive order explicitly rules out anything resembling mandatory licensing or preclearance. Both statements can be true at once, and that's exactly the problem. Nobody had to write a law. A cabinet secretary approving a customer list by name is a licensing regime with better PR.

This is the same week the European Commission stood up its own mandatory pre-market evaluation for advanced models, enforceable from August 2. Two governments, two legal theories — one calls itself voluntary, one calls itself law — arrived at the identical structure: a state actor deciding, ahead of the market, who gets to touch the frontier first. For a venture-backed startup building on top of GPT-5.6, whether it made Lutnick's list of twenty mattered more, for twelve days, than any amount of funding or technical merit.

The White House's denial is worth taking seriously on its own terms — no statute currently requires this review, and the administration is right that nothing forces the next model release through the same process. But precedent doesn't need force of law to compound. OpenAI complied once, under a request it could have declined. The next frontier lab now has to explain, to its own board and to the market, why it wouldn't.