OpenAI's Most Capable Model Yet Needed a Government Sign-Off First

OpenAI announced GPT-5.6 Sol, Terra, and Luna on June 26, 2026 — its most capable model family yet, and the first OpenAI release where a government reviewed the model before customers could. The initial rollout went to "a small group of trusted partners" through the API and Codex, after OpenAI previewed the models' plans and capabilities to the U.S. government and began the limited release, in the company's own words, "at the government's request."

Three tiers, one new gate

Sol is the flagship, built for the hardest problems: complex coding, computer use, and cybersecurity and biology research. Terra is a mid-tier model priced for high-volume business tasks — customer support, internal tools, document analysis — at roughly half of GPT-5.5's cost for comparable performance. Luna is the cheapest and fastest, aimed at summarization, drafting, and routine automation. Pricing runs $5/$30 per million input/output tokens for Sol, $2.50/$15 for Terra, and $1/$6 for Luna.

Two new reasoning controls ship alongside the models: max, which gives Sol extended time to reason through a problem, and ultra, which deploys multiple subagents on a task instead of one. GPT-5.6 Sol running in Ultra mode posted a new state-of-the-art score on Terminal-Bench at 91.9%, and Sol on its own set a new mark on Terminal-Bench 2.1 — both benchmarks measuring an agent's ability to actually operate a terminal environment, not just answer questions about one.

The review process is the actual story

What makes this launch different isn't the benchmark scores — it's that OpenAI didn't control the rollout timeline on its own. The Trump administration required advance visibility into Sol's capabilities before OpenAI could ship broadly, and OpenAI has said publicly it disagrees with this becoming "the long-term default," arguing that gating access this way restricts it for "users, developers, enterprises, cyber defenders" who have legitimate reasons to want the model now rather than in "the coming weeks," when OpenAI expects broader availability across ChatGPT, Codex, and the API.

The categories driving that review track closely with where Sol actually gained the most ground: coding, cybersecurity vulnerability research and exploitation, and genomics analysis. That's not a coincidence — it's the same dual-use logic Anthropic applied a few weeks earlier when it split Claude Fable 5 from the more capable, tightly restricted Claude Mythos 5: the capability gains that make a model genuinely more useful for defenders are close to identical to the ones that make it more dangerous in the wrong hands. Anthropic drew that line itself, through a specific partner program. Here, a government drew it for OpenAI, on OpenAI's own flagship release.

What to actually do with this

If you're planning work around Sol's coding or cybersecurity capabilities, don't architect anything on the assumption of near-term general access — "a small group of trusted partners" is a narrow list, and OpenAI hasn't named who's on it. Track the general-availability announcement rather than the preview, since pricing and rate limits could still shift before then. And treat this as a signal worth watching regardless of whether you use OpenAI or Anthropic: government pre-clearance for a frontier release, applied to a Western lab for the first time here, is exactly the kind of precedent that tends not to stay a one-off.