OpenAI unveils GPT-5.6 in three tiers — Sol, Terra, and Luna — with US government gatekeeping

OpenAI has launched a limited preview of GPT-5.6, a new model family split into three tiers — and access is deliberately narrow by design, at the request of the US government. Around 20 approved partners can use the models via API and Codex as of June 26, 2026. General availability, including ChatGPT access, is expected "in the coming weeks," per OpenAI official.

The three tiers

The lineup is tiered by capability and cost. Sol is the flagship — the most powerful of the three, with a new "max reasoning" mode that gives the model extra time to work through complex problems, plus an "ultra" multi-agent mode that spins up virtual sub-agents to handle tasks in parallel. OpenAI says Sol performs particularly well on biology-related tasks.

Terra is the balanced option, positioned to match GPT-5.5 performance at half the price: $2.50 per million input tokens and $15 per million output tokens, versus GPT-5.5's $5/$30. For teams already paying for GPT-5.5, that's a straightforward cost argument.

Luna is the cheapest and fastest, at $1 per million input tokens and $6 per million output tokens. At that price point, bulk processing tasks — summarization, classification, data extraction — become dramatically cheaper to run at scale.

Safety first, then speed

All three models arrive with a multilayer safety stack. OpenAI trained the models to resist jailbreak attempts and added a real-time filter that can pause generation mid-response if something suspicious is flagged. During the preview period, users may occasionally hit delays or false-positive blocks as a result.

To stress-test the safety systems before release, OpenAI used other AI models to run continuous adversarial attacks — hundreds of thousands of hours of probing — to surface weaknesses ahead of the public launch.

The phased rollout itself is a story. The US government requested the limited preview before a broader release, and TechCrunch confirmed OpenAI's statement that such restrictions "shouldn't be the norm" — signaling the company sees this as a one-off precaution rather than a new standard.

What to expect next

No specific date has been set for general availability, and the models are not yet accessible in ChatGPT for regular users. The API-only preview is currently limited to vetted partners. Once the full rollout begins, all three models will be available through the standard ChatGPT and API access points. For developers and businesses watching costs, Terra's pricing is the number to watch — at GPT-5.5 capability levels, it could reshape how teams budget for production AI workloads.