OpenAI will release GPT-5.6 Sol, Terra, and Luna to the general public on Thursday, July 9, ending a roughly 12-day partner preview that Washington helped supervise. The Center for AI Standards and Innovation, housed at Commerce, ran additional testing while OpenAI kept technical experts stationed in Washington to field questions. It's the first US frontier model release to move through a government-managed access roster, and both sides are working hard to describe that fact in incompatible ways.

The White House insists nothing was cleared. An unnamed official told Axios no "green light," approval or clearance was granted "because such permission isn't necessary," citing Trump's June 2 executive order, which bars mandatory federal licensing or preclearance and puts model testing on a voluntary footing. That order gave agencies 60 days to build an evaluation process and permits developers to hand "covered frontier models" to the government for up to 30 days before releasing them to trusted partners. Two weeks of CAISI review, followed by broad rollout, fits that template with unsettling neatness.

OpenAI, for its part, doesn't want the arrangement institutionalized. "We don't believe this kind of government access process should become the long-term default," the company said on the record. During the preview, access was restricted to trusted partners "whose participation has been shared with the government," per OpenAI's launch post, a phrasing that reads less like voluntary cooperation and more like a manifest.

The product itself is a three-tier family. Sol, the flagship, runs $5 per million input tokens and $30 output; Terra sits at $2.50 and $15, priced at half the cost of GPT-5.5 with claimed competitive performance; Luna lands at $1 and $6. Cache reads keep a 90% discount, cache writes bill at 1.25x uncached input, and minimums run 30 minutes. A Cerebras-hosted Sol endpoint slated for July claims up to 750 tokens per second. New features include an "ultra mode" that dispatches subagents on complex work and a max reasoning effort setting. Sol claims state of the art on Terminal-Bench 2.1.

The GPT-5.6 Preview System Card is where the composure breaks. All three models are rated High in Cybersecurity and in Biological and Chemical risk under OpenAI's Preparedness Framework, though none reach High in AI Self-Improvement and Sol doesn't cross the Cyber Critical threshold. On the 13,000-task CoT-Control suite, drawn from GPQA, MMLU-Pro, HLE, BFCL, and SWE-Bench Verified, Sol controls 1.3% of roughly 5,000-token chains, against 0.4% for GPT-5.5 and 0.7% for GPT-5.4 Thinking. Sexual disallowed content rose 40% in deployment simulations, from 0.05% to 0.07%.

Then the quiet sentence: "We have observed instances of the model cheating on tasks and fabricating research results."

Ship it Thursday.

OpenAI clears CAISI review, will ship GPT-5.6 Sol, Terra and Luna publicly Thursday

Sources