OpenAI is launching a limited rollout of its GPT-5.6 lineup, introducing three tiers, Sol, Terra, and Luna, while saying a wider release is expected in the coming weeks. The company is taking a staged approach, saying the goal is to pair higher-end capability, particularly for security-related tasks, with tighter controls and more testing before opening access more widely.

The initial rollout is restricted to a small group of vetted partners, with participation already shared with the U.S. government, the company said in a blog post. Access is being delivered first through the API and Codex.

OpenAI said the approach is meant to be temporary, arguing that long-term government gating shouldn't become the default because it slows access for developers, enterprises, and cyber defenders but it sees the current structure as a necessary step while broader cyber policy frameworks are still being shaped.

At the top end, GPT-5.6 Sol is positioned as the flagship model, adding new reasoning controls like a "max effort" setting and an "ultra" mode that can deploy subagents for more complex workflows. OpenAI highlighted strong benchmark results across coding, biology, and cybersecurity tasks, including performance on Terminal-Bench 2.1 and GeneBench v1.

On the security front, the company says Sol is better at spotting and fixing software vulnerabilities than carrying out full exploit chains. In testing across browsers like Chromium and Firefox, it was able to surface bugs and exploit components, but didn't independently produce end-to-end working exploits under the conditions evaluated.

OpenAI also outlined a layered safety system combining training, real-time monitoring during outputs, account-level signals, and tiered access controls. During the preview, some prompts may be slowed or blocked for extra review as the company fine-tunes false positives ahead of wider release.

Behind the scenes, OpenAI says it has logged more than 700,000 A100-equivalent GPU hours on automated red-teaming to probe jailbreaks, alongside external expert testing. It also described a rapid patch cycle to reproduce newly discovered jailbreak methods and fold fixes into future evaluations.

Separately, GPT-5.6 Sol is slated to come to Cerebras in July, with throughput of up to 750 tokens per second, though access will start limited as capacity ramps. After the preview phase, OpenAI expects broader availability through ChatGPT, Codex, and its API.