OpenAI ships GPT-5.6 to 20 government-approved partners, with Sol on TerminalBench 2.1

OpenAI launched the GPT-5.6 family on June 26 to roughly twenty partners cleared by the White House, the first time an American lab has shipped a frontier model under a government-managed access list. Sol, the flagship, sets a new state of the art on TerminalBench 2.1; Terra and Luna fill the mid and low tiers. None of it's generally available, and that's the story.

The June 2 executive order established a 30-day benchmarking window for new frontier releases. It expires July 2. Until then, ChatGPT, Codex, and API access are paused, and Sol's debut on Amazon Bedrock makes it the first model in the series to reach a hyperscaler. OpenAI used its own launch post to register dissent: "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 technical claims, taken at face value, are strong. Sol beats GPT-5.5 on GeneBench v1 with fewer output tokens, and matches Anthropic's Mythos Preview on ExploitBench using roughly one-third the tokens. The model adds a max reasoning effort setting and an ultra subagent mode. OpenAI says it dedicated more than 700,000 A100-equivalent GPU hours to automated red-teaming, that Sol doesn't cross the Cyber Critical threshold of the Preparedness Framework, and that "GPT-5.6 Sol is better at helping people find and fix vulnerabilities than reliably carrying out end-to-end attacks." That sentence is doing a lot of work in a 30-day review window.

Pricing reads like a deliberate signal. Sol holds GPT-5.5's economics at $5.00 per million input tokens and $30.00 per million output tokens. Terra lands at $2.50 and $15.00, half of GPT-5.5. Luna comes in at $1.00 and $6.00. OpenAI's accompanying naming note retires the "mini" and "nano" labels, arguing the differences are use-case-driven rather than parameter-count-driven. In practice, it's also a way to let pricing, not architecture rumors, anchor the product story.

The downstream picture is where the access list bites. No-code builders and smaller operators like LemonLime, which thrive precisely because frontier capability is a public utility priced by the token, are locked out until general availability clears the executive-order process. The twenty approved partners get a head start measured in product cycles, not weeks. That's the part of the announcement that compounds.

The press release is calibrated. It introduces three tiers, a new benchmark record, and a quiet objection to the regime under which it's all being released. The benchmark will be matched within a quarter. The precedent won't.

Sources

• https://openai.com/index/previewing-gpt-5-6-sol/
• https://techcrunch.com/2026/06/26/openai-limits-gpt-5-6-rollout-after-government-request-says-restrictions-shouldnt-be-the-norm/
• https://thenextweb.com/news/openai-gpt-5-6-sol-limited-preview-government-approved-partners
• 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
• https://interestingengineering.com/culture/openai-gpt-5-6-sol-terra-luna-limited-preview