OpenAI previews GPT-5.6 Sol

OpenAI previewed its next model generation, a three-model GPT-5.6 family: Sol, the flagship, with “max” and “ultra” reasoning modes and gains OpenAI described in coding, biology and cybersecurity; Terra, a mid-tier model matching GPT-5.5 at roughly half its cost; and Luna, a cheaper, faster variant. OpenAI called Sol its strongest cybersecurity model yet, framed around defensive work such as vulnerability discovery, code review and patching, backed by what it described as its most robust safety stack to date — layered training-time and inference-time protections against misuse of the model’s offensive capability.

Access, however, was not open. Under a request from the Trump administration tied to the frontier-AI executive order signed earlier in June, OpenAI limited the initial preview to a small group of trusted partners it said it had vetted with the government, rather than shipping the models broadly in ChatGPT, Codex and the API as it would for an ordinary release. OpenAI pushed back on the arrangement in its own announcement, writing that it did “not believe this kind of government access process should become the long-term default,” and said a wider public rollout would follow within weeks pending further testing. GPT-5.6 reached general release in July.