OpenAI Limits GPT-5.6 Sol Launch to Government-Approved Partners
OpenAI has begun a limited preview of GPT-5.6, its newest model generation, but the rollout comes with an unusual catch: access is restricted to roughly 20 companies that the U.S. government has signed off on. The company says it agreed to the restriction at the request of the Trump administration, which raised national security concerns tied to the model's advanced cybersecurity abilities.
The release, which started June 26, marks one of the clearest signs yet that Washington now treats the most capable American AI systems as products that warrant government review before they reach the open market. For founders and operators who build on frontier models, it is a preview of a new gating layer between a model's announcement and the day they can actually use it.
What OpenAI announced
GPT-5.6 arrives in three variants that double as a new naming system. Going forward, the version number signals the model generation while the names mark permanent capability tiers. Sol is the flagship, built for high-intensity reasoning and described by OpenAI as its strongest model to date. Terra is the everyday workhorse, with performance the company says is competitive with the prior GPT-5.5 at roughly half the operating cost. Luna is the fast, low-cost option at the bottom of the lineup.
The generation also introduces a new "max" reasoning effort that gives Sol more time to think through hard problems, plus an "ultra" mode that splits a task across multiple subagents to speed up complex work. Pricing undercuts OpenAI's previous top tier. Sol runs $5 per million input tokens and $30 per million output tokens, Terra is priced at $2.50 and $15, and Luna sits at $1 and $6. For comparison, the earlier flagship cost $10 and $50 for the same volumes, so the new lineup pushes more capability toward a lower price point.
For now the models are reachable only through the API and OpenAI's Codex tool, and only by approved partners. OpenAI says it expects to widen access to more companies within a week and to reach a broad release across ChatGPT, Codex, and the general API in the coming weeks.
Why the government is involved
The limited rollout traces back to an AI cybersecurity executive order President Trump signed earlier in June. That order asks AI developers to submit their most powerful models for voluntary government review 30 days before a public launch. OpenAI says it spent roughly a month previewing GPT-5.6 with officials, including meetings between chief executive Sam Altman and the White House in early June.
OpenAI made plain it views the arrangement as temporary and far from ideal. "We don't believe this kind of government access process should become the long-term default," the company wrote, arguing that prolonged gatekeeping keeps useful tools away from developers, enterprises, and cyber defenders who need them. Even so, it framed cooperation as the fastest route to a wide release while a more durable framework gets worked out.
OpenAI is not alone. The restrictions follow a similar episode with Anthropic, which suspended access to its Fable 5 and Mythos 5 models after a government directive, then began restoring access once it received permission to redeploy Mythos to a select group of organizations. According to reporting cited by multiple outlets, most leading developers, including Google, xAI, and Microsoft, have been giving the government early looks at their newest models. Meta has been described as the lone holdout, with officials reportedly pressing it to participate.
The cybersecurity question at the center
The sharpest concern around GPT-5.6 is its security capability. OpenAI positions Sol as its most capable model yet for cybersecurity work, and it leans hard on the distinction between defense and offense. The company says Sol is better at helping people find and fix vulnerabilities than at reliably carrying out full attacks, and that its capabilities do not cross the "critical" threshold defined in OpenAI's preparedness framework.
Independent-style benchmarks back the framing. On a command-line coding test, Sol set a new state of the art. In security evaluations, OpenAI reported that Sol matched a rival system while using about a third of the output tokens. During tests against the Chromium and Firefox codebases, the model isolated bugs and basic exploitation building blocks but did not independently assemble a working full-chain exploit.
To manage the dual-use risk, where defensive research can look a lot like offensive activity, OpenAI built a layered safety stack. Beyond standard refusals learned during training, the system runs real-time classifiers on biology and cybersecurity inputs. When something looks anomalous, output pauses while a second reasoning model reviews the conversation, and some cases trigger account-level checks meant to separate legitimate researchers from bad actors. The company says it spent more than 700,000 GPU hours on automated red-teaming aimed at finding universal jailbreaks rather than one-off failures, and it pledged a rapid-response process to patch new ones.
What it means for founders and operators
For startups and operators, the immediate takeaway is timing. A model can be announced, benchmarked, and priced, and still be out of reach for weeks while a clearance process plays out. Teams that plan product launches around a specific model's availability now have a new variable to manage, and "generally available soon" carries more uncertainty than it used to.
The pricing shift is the more encouraging signal. Cheaper inference on a more capable lineup lowers the cost of running AI features at scale, and the Terra tier in particular targets the everyday workloads that most software companies actually ship. Builders weighing margins on AI-heavy products may find the math improves.
There is also a strategic wrinkle worth watching. As frontier models gain real security muscle, the same capabilities that help a defender harden code can, in the wrong hands, accelerate an attack. Operators should expect more friction, more identity and use-case verification, and more guardrails baked into the tools they rely on. By August, the executive order directs the administration to stand up a classified process for assessing AI cyber capabilities and to define which systems count as "covered frontier models," a label that could shape how the next wave of releases reaches the market.
Frequently asked questions
What is GPT-5.6 Sol?
Sol is the flagship variant of OpenAI's GPT-5.6 model generation, built for high-intensity reasoning and described by the company as its strongest and most capable model for cybersecurity tasks to date.
Why is access limited to about 20 companies?
OpenAI agreed to a restricted rollout at the request of the Trump administration, which cited national security concerns. Participation of the initial partners was shared with the government as part of a review tied to a June executive order.
How do Sol, Terra, and Luna differ?
Sol is the top reasoning model, Terra is the everyday tier with performance OpenAI says rivals GPT-5.5 at about half the cost, and Luna is the fastest and cheapest option. The names are meant to be permanent capability tiers across future generations.
How much does GPT-5.6 cost?
Sol is priced at $5 per million input tokens and $30 per million output tokens. Terra runs $2.50 and $15, and Luna costs $1 and $6, undercutting OpenAI's previous flagship pricing of $10 and $50.
When will GPT-5.6 be widely available?
OpenAI says it expects to expand access to more companies within about a week and to reach a broad release across ChatGPT, Codex, and the general API in the coming weeks, pending continued testing and government sign-off.