OpenAI will publicly launch all three GPT-5.6 variants on Thursday, July 9, ending a two-week window in which the model family was restricted to a small group of trusted partners Engadget. The announcement came via OpenAI's account on X, confirming reporting from Axios that the U.S. Department of Commerce had approved a broad rollout Axios via Yahoo Finance.

The series comprises three models: Sol, Terra, and Luna, first previewed to the developer community on June 26 OpenAI Community. Sol is billed as OpenAI's flagship for developers and enterprises, described in the company's own preview post as delivering state-of-the-art performance in reasoning, coding, and science, with strengthened cybersecurity capability and OpenAI's most advanced safety stack applied to it OpenAI. It is priced at $5 per million input tokens and $30 per million output tokens, OpenAI's highest rate to date. Terra targets everyday use cases, promising performance comparable to GPT-5.5 at roughly half the cost — $2.50 per million input tokens and $15 per million output tokens. Luna is the budget tier, at $1 per million input tokens and $6 per million output tokens Engadget.

The delay between preview and general availability traces back to a Trump administration request. In early June, the president signed a scaled-back AI cybersecurity order asking companies to voluntarily submit their most powerful models to government review 30 days ahead of public release Engadget. OpenAI complied by limiting GPT-5.6's late-June debut to trusted partners rather than the general public, even as it published a system card on its Deployment Safety Hub and a help center article describing the preview help.openai.com. At the time, OpenAI stated publicly that such restrictions "shouldn't be the norm" TechCrunch.

The review itself ran through the Department of Commerce's Center for AI Standards and Innovation, which put GPT-5.6 through additional testing. OpenAI sent technical staff to Washington to address the agency's questions directly Engadget. That process, plus further meetings between the company and administration officials, produced the clearance now allowing Thursday's wider release.

OpenAI is not the only lab to have run into this review mechanism. Anthropic previously blocked all access to its Mythos cybersecurity model and its Fable model after being ordered to bar foreign nationals from using them. The company later regained permission to redeploy Mythos 5, and eventually Fable 5, once government concerns were addressed Engadget.

The pattern across both companies is consistent: a frontier model is built, quietly shown to a narrow set of partners or held back from certain user classes, subjected to federal review, and then cleared for wider release once officials are satisfied. It is a materially different posture from the voluntary, industry-led disclosure norms that governed prior model launches, where companies set their own release cadence and consulted government largely on request rather than under a standing 30-day review clause.

Whether this becomes a durable feature of the U.S. regulatory landscape or a transitional arrangement tied to one administration's cybersecurity order is a separate question from what happened this week. OpenAI's own language — that such restrictions "shouldn't be the norm" — signals the company views the current arrangement as exceptional rather than a template it welcomes long-term. Worth flagging: that view was published while the company was still complying with the order, which suggests the friction is being managed rather than resolved.

For enterprise buyers, the more immediate story may be pricing. Terra's positioning — GPT-5.5-class performance at half the cost — extends a trend visible across the last several model generations, where each new release pushes down the price of "good enough" capability while a smaller, pricier flagship absorbs the frontier gains. Sol's $30 per million output token rate sits well above Terra and Luna, reserved presumably for workloads where the cybersecurity and reasoning improvements justify the premium. Organizations running cost-sensitive inference at scale will likely gravitate toward Luna or Terra by default, reserving Sol for the narrower set of tasks where its stated strengths in coding and science actually move the needle.

In this author's view, the more consequential thread here is not the pricing tiers but the review mechanism itself. A 30-day government preview window for frontier models, applied twice now across two major labs within a matter of weeks, is the kind of infrastructure that tends to outlast the specific circumstances that created it. Whether that proves durable will depend less on this week's clearance than on how the next model cycle handles the same request.