OpenAI previewed its GPT-5.6 series on Friday behind a launch list of roughly 20 companies, each individually approved by the U.S. government, a constraint the company says it accepted “at the behest of the U.S. government” while also arguing the arrangement shouldn’t stick.

The three models, Sol, Terra, and Luna, are priced as a conventional tiered stack: $5 input and $30 output per million tokens at the top, $2.50 and $15 in the middle, $1 and $6 at the bottom. Terra reportedly matches GPT-5.5 at half the cost. A Cerebras deployment of Sol is planned for July at up to 750 tokens per second. By the company’s own framing, Sol is its strongest model yet, with agentic chops in coding, biology, and cybersecurity, a new “max” reasoning mode, and an “ultra” mode coordinating subagents. OpenAI says Sol is better at fixing vulnerabilities than running end-to-end attacks, and still doesn’t cross its “critical” cybersecurity threshold.

The structural story isn’t the spec sheet. It’s the guest list.

For a month before launch, OpenAI had been previewing GPT-5.6 with the government, including meetings Sam Altman took at the White House in early June. According to Bloomberg and TechCrunch, the company didn’t anticipate that the previews would graduate into per-customer approval authority. Then they did. Axios reported the ~20-partner cap; broader access is expected “next week,” with full release in “coming weeks.”

OpenAI’s own statement reads like a hostage note dressed as a press release: “we don’t believe this kind of government access process should become the long-term default. It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them.”

The context cuts against the protest. Anthropic recently disabled two of its latest models to comply with a Trump administration export-control directive, and is negotiating safeguards before releasing another. An Executive Order requires the administration to stand up a classified process by August to assess AI models’ cyber capabilities and decide which qualify as “covered frontier models.” The August deadline is the actual launch date being optimized around. June’s rollout is a dry run for the regime that arrives in eight weeks.

What’s notable isn’t that a frontier lab complied. It’s that the compliance came with on-the-record dissent, which is itself a form of narrative management: shipping the precedent while disowning it.

Sources