OpenAI gates GPT-5.6 Sol access to Trump-approved customers over cyber concerns

OpenAI is not rolling out GPT-5.6 Sol to everyone. The company said on Friday it will stagger the rollout of its newest and most powerful AI model after a request from the Trump administration, with access limited first to customers cleared by the U.S. government.

The model is called GPT-5.6 Sol, and OpenAI says it hopes to make Sol, along with its expanded tier companions Terra and Luna, generally available in the coming weeks. In other words: the “most powerful” tier is real, but the market does not get it all at once.

Here is what OpenAI is offering in the meantime. Sol sits at the top of a new tier of more advanced models, alongside Terra (more efficient) and Luna (cheaper). OpenAI describes Sol as its strongest model yet, saying it can complete 50% of long-running professional tasks and tops all previous OpenAI models on coding capabilities. It is also emphasizing cybersecurity as a core differentiator. The company says Sol made its strongest gains in cybersecurity, specifically vulnerability and exploitation.

Access itself is being used like a gate. OpenAI frames the arrangement as “ongoing engagement” where it previewed plans and capabilities ahead of launch, and then, “at their request,” it is starting with a limited preview for a small group of trusted partners whose participation has been shared with the government. The list expands next week. The company also said it is not in favor of this process becoming a “long-term default,” calling it a short-term step while it works with the administration to develop a cyber executive order framework and a “repeatable process for future model releases.”

What does that mean in practice for security posture and product rollout? OpenAI is pairing the gated preview with “its most extensive safeguards to date.” The company says the preview will police its own use. For higher-risk cases, it says a larger model will review the conversation and could withhold a response if it is judged to violate policy.

On the technical risk side, OpenAI says Sol did not cross the “Cyber Critical” threshold in its “Preparedness Framework.” It describes tests with Firefox and Chrome where it found the seeds of an exploit but did not produce a working one. OpenAI also says it spent 700,000 GPU hours hacking itself to identify vulnerabilities. Human testers will conduct two more weeks of tests before launch. In other words, the gating is not just political. OpenAI is tying it to measurable cybersecurity checks, benchmarks, and pre-launch red-teaming.

If you are tracking the regulatory pattern, this is the second time in a month a frontier lab’s most powerful model has been held back from general release over fears about cyber capabilities. In early June, the Commerce Department issued export controls on Anthropic that forced the lab to cut off foreign access to two top models, citing national security concerns. Anthropic disputed the order but pulled the models offline because it had no choice.

Earlier this month, President Trump signed an executive order directing federal agencies to establish a framework under which AI companies could voluntarily provide the government with early access to powerful new models for up to 30 days before broader release. OpenAI describes its situation as voluntary, unlike Anthropic’s export-control driven shutdown. Still, critics argue the end result can look like an improvised licensing regime rather than a coherent regulatory system with consistent rules and boundaries.

That tension matters because it changes how the “frontier AI” business gets priced. Sol is priced at $5 per million input tokens and $30 per million output tokens, compared with Terra at $2.50 and $15 and Luna at $1 and $6. But pricing is not the only constraint. If customers need government approval to access the model, procurement timelines, go-to-market sequencing, and revenue recognition all become entangled with government workflows.

OpenAI also introduced two new modes, “max” and “ultra,” which allow the model to reason longer and coordinate agents for specific tasks. On a key cybersecurity benchmark, OpenAI previously said Sol was “competitive with” Anthropic’s Mythos. It also says Sol uses approximately one-third of the tokens used by Mythos but appears to lag slightly behind Mythos 5. That comparison matters because it signals that the cyber performance is not theoretical. It is directly tied to measurable capability and therefore directly tied to why Washington cares.

There is a strategic stake here for every AI operator, investor, and board member: access could become conditional even when models are “almost ready.” Capability concerns, export controls, and 30-day early-access frameworks create a world where launch timing is negotiated, not purely engineered. The administration rescinded a Biden-era requirement on the first day in office that asked AI companies to submit safety tests to the government, calling it overly burdensome. Yet what is emerging, as critics put it, is an ad hoc approach that repurposes existing legal authorities into something closer to backdoor licensing, with unclear criteria and limited appeal paths.

OpenAI says Sol is linked to the June 2 executive order, which directed agencies to build a framework for vetting models before release. Since that framework does not exist yet, OpenAI says it conducted a phased rollout at the government’s request. For executives at other frontier labs, the takeaway is simple but not comfortable: if you are building the most capable cyber-adjacent systems, you should assume rollout can be delayed, segmented, or redirected by security gatekeeping until the government’s process is fully operational.