OpenAI shipped GPT-5.6 Sol, Terra, and Luna to general availability on July 9 across ChatGPT, Codex, and the API, closing a 13-day preview window that Bloomberg reports had been "initially delayed by the Trump administration" and which VentureBeat says was scoped to roughly 20 organizations at the government's request. The models are the story on paper. The release process is the story in practice.

Sol is priced at $5 per million input tokens and $30 per million output tokens, with Terra at half of Sol and Luna at $1 in, $6 out. OpenAI's self-reported benchmarks put Sol at a state-of-the-art Terminal-Bench 2.1 score and claim it uses approximately one-third of the output tokens of Mythos Preview on ExploitBench. The company describes the lineup as "durable capability tiers that can advance on their own cadence," and says Sol was "intentionally optimized to favor defensive cybersecurity work over offensive exploits." A Cerebras deal targets up to 750 tokens per second later this month.

The safety story is denser. VentureBeat reports OpenAI spent roughly 700,000 A100e GPU-hours on automated red-teaming aimed at "universal jailbreaks," with external evaluations from the U.K. AISI, SecureBio, and Irregular cited in the system card. OpenAI's own late-June post on the arrangement was blunt: "we don't believe this kind of government access process should become the long-term default." The company says it "previewed our plans and the models' capabilities" to the government before launch.

That process now has a paper trail. A June 2 Trump executive order gave federal agencies 30 days to design a benchmarking process for frontier models, and six cabinet agencies have been told to converge on a final procedure by early August. What actually happened during the 13-day preview, and against what criteria roughly 20 organizations were selected, hasn't been disclosed.

The critics are worth quoting directly. Sriram Krishnan, until last month a White House senior AI adviser, told the Financial Times "there will not be an FDA for AI." Mina Narayanan of Georgetown CSET told TechCrunch: "I don't have enough information to say whether they're adequate or not." Andy Konwinski was sharper: "It's existentially a problem. Safety or not, it's about who has the power to make decisions."

Sol arrives alongside ChatGPT Work, a long-running agent OpenAI also introduced July 9. The desk's read: the model is genuinely strong, and Terra and Luna extend the price ladder in useful directions. But the interesting artifact isn't the weights. It's that a private company and a handful of cabinet agencies now run a bespoke, undocumented pre-release review, and both sides are on record saying they'd prefer it didn't calcify. It probably will anyway. That's how the TARP framing of 2008 became the template for every subsequent crisis-era Treasury intervention: emergency procedures acquire tenure.

Sources

https://openai.com/index/previewing-gpt-5-6-sol/
https://techcrunch.com/2026/07/09/how-did-the-government-decide-openais-frontier-model-was-safe-to-release/
https://techcrunch.com/2026/06/26/openai-limits-gpt-5-6-rollout-after-government-request-says-restrictions-shouldnt-be-the-norm/
https://www.bloomberg.com/news/articles/2026-07-09/openai-unveils-chatgpt-work-agent-to-field-tasks-for-hours
https://venturebeat.com/technology/openai-unveils-gpt-5-6-sol-terra-and-luna-models-but-only-accessible-to-limited-preview-partners-for-now-per-us-gov