OpenAI began a limited preview of its GPT-5.6 model family on June 26, 2026, led by a flagship called Sol and joined by two cheaper tiers, Terra and Luna. The catch: almost nobody can use it yet. At the request of the U.S. government, OpenAI is restricting access to a small group of vetted partners through the API and Codex, with no ChatGPT access during the preview.
The government part is the actual story
Buried in a launch about coding and cybersecurity scores is a genuinely odd admission. OpenAI says it previewed the models and its release plans with the government beforehand, and at the government's request is starting narrow, with partner names shared with officials. Then, in the same launch announcement, the company pushes back on the very arrangement it's complying with, saying this kind of access process shouldn't become the long-term default.
"We don't believe this kind of government access process should become the long-term default." Which is a strange thing to write in a post announcing that you've done exactly that, even framed as a short-term step toward broader availability.
This follows a June 2 executive order setting up a pre-launch benchmarking process for frontier models. Anthropic got the harsher version of this treatment earlier, when access to two of its models was pulled. OpenAI is, for now, choosing to comply loudly and complain in writing.
What's new under the hood
Two additions stand out. Sol gets a new max reasoning effort, giving it the longest thinking budget yet, and a new ultra mode that spins up subagents to handle work beyond what a single agent manages. OpenAI doesn't say how ultra interacts with existing pro-tier offerings, and the pricing for running it isn't broken out in the launch post, so treat the cost as an open question.
The naming is also new. The number marks the generation, while Sol, Terra, and Luna mark capability tiers that can move on their own schedules. Whether that's clearer than a version number is something users will decide on their own.
The numbers, with a grain of salt
On cybersecurity, OpenAI claims Sol matches its Mythos Preview competitor on ExploitBench while using roughly a third of the output tokens, and that all three models improve on ExploitGym, a UC Berkeley benchmark built with OpenAI and other labs. These are self-reported, and OpenAI says it'll publish a fuller evaluation suite at general availability, so the picture is partial by the company's own admission.
The framing OpenAI leans on hard: Sol is better at finding and fixing vulnerabilities than running attacks end to end, and it doesn't cross the company's Cyber Critical threshold. In Chromium and Firefox tests, it found bugs and exploitation primitives but didn't autonomously build a full working exploit under the conditions tested. The hedge is right there in the post, that benchmark thresholds can't capture every way a model gets combined with other tools.
Pricing and the speed flex
Sol runs $5 per million input tokens and $30 output. Terra is $2.50 and $15, pitched as GPT-5.5-class performance at half the cost. Luna sits at $1 and $6. Prompt caching changes too: cache writes now bill at 1.25x the uncached input rate, while cache reads keep the 90% discount.
The headline number for the speed obsessed is Cerebras. OpenAI plans to run Sol on Cerebras hardware at up to 750 tokens per second starting in July, limited to select customers first as capacity expands. That's the kind of figure that's either a serious shift in how agentic workloads feel or a launch-day estimate that softens in production. We'll find out when it's not gated to a handful of accounts.
General availability across ChatGPT, Codex, and the API is promised in the coming weeks, with no firm date. The Cerebras rollout starts in July.
