Labs

OpenAI's GPT-5.6 ships
behind a government
gate

The Sol, Terra, and Luna models debuted Friday in a limited preview restricted to a small group of partners whose names were cleared with the Trump administration — the first time Washington has staged a flagship model launch.

By Casey Vanderhoof Frontier labs correspondent · June 28, 2026

OpenAI unveiled GPT-5.6 on Friday, with the flagship Sol model and its smaller siblings Terra and Luna restricted to a preview audience whose names, per Bloomberg, were “approved by the US government.” The company disclosed the arrangement itself, in a blog post that framed the gating as a “short-term step” while signaling broader availability “in the coming weeks” through the API and Codex.

That disclosure is the story. A lab announcing a frontier model has, until now, been a commercial event. Friday’s launch was a regulatory one.

The lineup follows the now-familiar three-tier ladder. Sol lists at $5 per million input tokens and $30 per million output tokens. Terra runs at half those rates, $2.50 in and $15 out, and OpenAI claims it matches GPT-5.5 at half the cost. Luna comes in at $1 and $6. The series introduces a “max” reasoning effort setting and an “ultra” mode that coordinates subagents on complex tasks. OpenAI says Sol edges Anthropic’s Claude Mythos 5 on coding workflows while using roughly a third of the output tokens.

The political mechanics matter more than the spec sheet. Reuters frames the staggered launch as a deferral tied to Washington’s push for early access to frontier systems to flag cyber and military misuse before wide deployment. CNBC notes OpenAI’s own claim that Sol, while its most capable cybersecurity model, is better at helping users fix vulnerabilities than executing end-to-end attacks and doesn’t cross the company’s “critical” threshold, defined as “unprecedented new pathways to severe harm.” The model cleared OpenAI’s internal bar and was still gated.

OpenAI is candid about the discomfort. It does “not believe this kind of government access process should become the long-term default,” and says it’s working with the administration toward “a repeatable process for future model releases” inside a cyber executive order framework. The contradiction is doing a lot of work in one sentence: build the pipes, then argue the pipes shouldn’t exist.

Dean Ball, the former White House AI adviser who’s also an incoming OpenAI employee, put it more sharply. A recent executive order asking AI companies to voluntarily submit their most advanced models for review up to 30 days before release has, he warns, produced a “de facto involuntary licensing regime” with open-ended delay risk. Anthropic offers the cautionary case. Per TechCrunch, Claude Mythos 5 was “effectively banned” by the administration this month, and CNBC reports the company disabled access to two of its latest models to comply with an export-control directive.

What looked in 2023 like a debate over voluntary safety commitments has settled, by mid-2026, into something closer to pre-clearance. The lab still ships. Washington decides who sees it first.

Sources