GPT-5.6 goes public after weeks of government limbo
OpenAI shipped GPT-5.6 to the public today. Three models launched together: Sol (the flagship), Terra (balanced for everyday work), and Luna (fast and cheap). The rollout comes after a delay of roughly two weeks triggered by US government scrutiny over national security concerns.
The delay started when the Trump administration asked OpenAI to hold back public access and limit initial release to a small group of trusted partners. OpenAI spent the interim conducting additional testing and meeting with federal officials from the Department of Commerce’s Center for AI Standards and Innovation.
But here is the twist. The White House pushed back hard today on the idea that OpenAI ever needed government approval.
“The Trump administration did NOT give OpenAI a ‘green light,’ approval, or clearance to release its models,” a White House spokesperson told Gizmodo. “No such permission is required or granted. The administration does not provide approvals for private companies to release AI models. Decisions on timing and scope of releases rest entirely with the companies.”
That statement directly contradicts an Axios report claiming the public release followed a government “green light.” Trump’s June 2 executive order established a voluntary safety testing framework where companies share models with federal officials for 30 days before public release. But the order explicitly states it does not create a mandatory licensing, preclearance, or permitting requirement. Cooperation was voluntary. The government never had a legal lever.
So the two-week delay was OpenAI choosing to play along, not the government forcing a halt.
The model itself ships real upgrades. GPT-5.6 introduces a new naming scheme: the number identifies the generation, while Sol, Terra, and Luna represent capability tiers that advance on their own cadence. Terra is competitive with GPT-5.5 at half the cost. Luna is the cheapest tier yet.
On benchmarks, Sol sets a new state of the art on Terminal-Bench 2.1, which tests command-line workflows requiring planning, tool coordination, and iteration. It also shows broad progress on GeneBench v1, a genomic and quantitative biology benchmark.
The cybersecurity capabilities are what triggered the government scrutiny in the first place. OpenAI calls GPT-5.6 Sol its most capable model yet for security work. On ExploitBench, it matches Anthropic’s Mythos Preview while using roughly a third of the output tokens. On ExploitGym, a benchmark created by UC Berkeley researchers in collaboration with OpenAI and other frontier labs, all three GPT-5.6 tiers show solid progress in offensive cyber capabilities as reasoning effort increases.
This is the same category of capability that got Anthropic’s Mythos 5 pulled offline last month. The Department of Commerce ordered Anthropic to block all foreign nationals from accessing Fable 5 and Mythos 5. Anthropic’s own internal testing later showed the capabilities that triggered the export ban could also be demonstrated by GPT-5.5 and earlier Claude models. In other words, the threat threshold for government intervention was already blurry.
Sol ships with two new reasoning modes. The first, max, gives the model more time to work through hard problems. The second, ultra, goes beyond a single agent by spinning up cooperating sub-agents to accelerate complex work. That sub-agent architecture is the same approach OpenAI previously teased for Codex.
The company says it spent weeks running automated red teaming and stress-testing the protection stack against real-world attacks. The layered defenses are designed to resist adversarial pressure while preserving legitimate access for code review, vulnerability research, and defensive security testing.
OpenAI also acknowledged the government preview process should not become a permanent fixture. The company framed the limited release as a short-term measure to reach broader availability while working with the administration on a repeatable framework for future launches.
GPT-5.6 Sol, Terra, and Luna are rolling out gradually starting today. OpenAI expects general availability in the coming weeks.