OpenAI began a limited preview of its next model family, GPT-5.6, on June 26, releasing three versions to a small group of partners rather than the public. The lineup splits into capability tiers:
- Sol, the flagship and OpenAI’s most powerful model yet;
- Terra, a balanced everyday option;
- and Luna, a fast, low-cost model.
Under a new naming scheme, the number marks the generation while the names mark durable tiers that can advance separately. The models are available through OpenAI’s API and its Codex coding tool to roughly 20 organizations whose participation was shared with the US government, with broader availability promised in the coming weeks.
The restricted rollout is the headline. CNN reported that the White House asked OpenAI to limit the release to government-approved partners because of the models’ advanced capabilities, and OpenAI confirmed it is starting small at the government’s request.
The step follows a June 2 executive order directing federal agencies to build a process for assessing powerful AI models before launch. OpenAI pushed back even as it complied, saying such a government access process “should not become the long-term default” because it keeps capable tools from defenders and developers who need them. The company said it sees the limited preview as the fastest path to a broad release while it works with the administration on a repeatable framework.
Sol’s gains are concentrated in agentic coding, biology and cybersecurity, and OpenAI added a deeper “max” reasoning setting and an “ultra” mode that splits work across subagents. In OpenAI’s own benchmarks, Sol leads the Terminal-Bench 2.1 coding test, scoring above GPT-5.5 and ahead of Anthropic’s Claude models and Google’s Gemini, though these figures are self-reported and unverified.
On safety, OpenAI rated all three models High for cybersecurity and biological risk under its Preparedness Framework, the first time its smaller tiers earned that label, but said Sol does not reach the “critical” cyber threshold; in tests it found bugs in Chromium and Firefox but did not autonomously build a full working exploit. OpenAI says it spent 700,000 GPU-hours on automated red-teaming and layered multiple safeguards, while warning the preview may sometimes block legitimate security work.
Why the Government Is Involved
The concern is dual-use. The same model that helps a defender find and patch a vulnerability could, in the wrong hands, help an attacker exploit one, and OpenAI says Sol is better at finding and fixing flaws than at carrying out end-to-end attacks.
Pricing signals where OpenAI is aiming: Sol costs $5 per million input tokens and $30 per output, the same as GPT-5.5 and about half the price of Anthropic’s Claude Fable 5, with Terra and Luna far cheaper.
OpenAI also plans to run Sol on Cerebras hardware at up to 750 tokens per second in July. The push is to get powerful, cheaper cyber and coding tools to enterprises quickly, which is exactly what makes the government cautious.
An Industry Under Watch
The episode shows a new normal taking shape. OpenAI’s constrained launch follows the government’s far blunter move against Anthropic this month, when officials effectively forced it to pull its Claude Fable 5 and Mythos 5 models over a jailbreak concern.
That OpenAI is now navigating a similar, if lighter, process suggests Anthropic was not being singled out, and that pre-release government review is becoming a feature of frontier AI in the US. For now the rules are improvised and case-by-case, decided model by model while the promised executive-order framework is still being written. How that framework lands will shape who gets access to the most capable AI, and how fast.
