OpenAI Previewed GPT-5.6 Sol, Terra, and Luna on June 26. Almost Nobody Can Use Them Yet.

On June 26, OpenAI announced GPT-5.6: a three-model family called Sol, Terra, and Luna. The catch is that access is currently restricted to roughly 20 vetted organizations, at the request of the US Office of the National Cyber Director and the Office of Science and Technology Policy. OpenAI says broader availability through ChatGPT, the API, and Codex is coming within weeks.

The Three Models

Sol is the flagship. OpenAI describes it as their strongest model yet for complex reasoning, long coding sessions, agent workflows, and cybersecurity tasks. It sets a new benchmark on Terminal-Bench 2.1, the standard evaluation for command-line and coding performance. Sol also introduces “max” and “ultra” reasoning modes: max runs deeper analysis, ultra spins up subagents for particularly complex work.

Terra is designed for everyday production workloads. It matches GPT-5.5 performance but at about half the cost, which makes it the practical default for most teams once it’s generally available.

Luna is the fast and cheap option for high-volume tasks where reasoning depth matters less than speed and price.

Pricing per million tokens:

Model	Input	Output
Sol	$5	$30
Terra	$2.50	$15
Luna	$1	$6

Prompt caching discounts apply across all three.

Why It’s Restricted

The government request isn’t surprising in context. Anthropic went through something similar with Mythos 5 and Fable 5, which faced non-American access restrictions. Sol’s capabilities in cybersecurity and biology apparently triggered the same concerns.

OpenAI noted that Sol was able to match competing models on ExploitBench, a cybersecurity benchmark, while using roughly one-third fewer output tokens. The company said Sol did not cross the “Cyber Critical” threshold and could not autonomously produce complete working exploits. Still, the government asked for a controlled rollout.

Safety measures built into the model include trained refusals, real-time output classifiers, and over 700,000 GPU hours of automated red-teaming before release.

What This Means for Codex Users

Sol is available through Codex today for the small group of preview partners. The Terminal-Bench 2.1 result is particularly relevant for Codex users since that benchmark tests complex command-line workflows, which is where Codex agents do most of their work. A Cerebras deployment is also coming in July 2026, rated at up to 750 tokens per second.

For everyone else, OpenAI hasn’t set a firm date. The phrase used was “coming weeks,” which based on their recent track record probably means July.

---

Sources: OpenAI announcement, OpenAI community forum, The AI Insider, Thurrott