OpenAI made GPT-5.6 Sol, Terra, and Luna generally available on ChatGPT, Codex, and the API on Thursday, ending a roughly 12-day window during which the models were restricted to partners “whose participation has been shared with the government.” The gating wasn’t a bug in the launch; it was the launch.

The June preview was the first live test of Trump’s AI executive order, signed earlier that month, which asks frontier developers to voluntarily hand cutting-edge models to federal agencies for capability assessment and gives those agencies 60 days to stand up an evaluation process. OpenAI complied, then publicly complained that the arrangement “keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them.” The company is now, per its own blog, helping the government construct a “repeatable process for future model releases.”

That phrase is the story. A voluntary framework tested once tends to become the default the second time.

Anthropic drew the harder end of the same regime. Late last month, the company had to disable access to Claude Fable 5 and Mythos 5 to comply with a Commerce Department export directive that was subsequently lifted. Two labs, two different pressure points, one architecture forming around them.

Sam Altman took the release lap on CNBC with Julia Boorstin, praising the technical chops of Secretary Lutnick, Secretary Bessent, and Director Karen Cross as surprisingly sharp, and arguing that any durable review regime ought to be global, quick, and not disproportionately tilted toward the U.S. The subtext is straightforward: the incumbent wants the process codified before a less friendly administration inherits it.

The product story is real too. Sol is priced at $5 per million input tokens and $30 output, and OpenAI claims it’s 54% more token-efficient than its predecessor on agentic coding. Terra ($2.50/$15) is pitched as matching GPT-5.5 at half the price. Luna comes in at $1/$6. A Cerebras-hosted Sol at up to 750 tokens per second lands this month for select customers, with a safety stack OpenAI describes as its most robust to date.

Twelve days of Washington review, and the release calendar barely blinked. That’s the point.

Sources