OpenAI just announced three new AI models at once, and if you’ve been following the company’s rollout pattern, you’ll know this launch is different from basically everything before it. The company is previewing its GPT-5.6 family, which comes in three variants: Sol, Terra, and Luna. These aren’t just minor upgrades—they’re a genuinely next-generation model family with stronger capabilities in coding, science, and cybersecurity, all paired with what OpenAI calls its most advanced safety stack to date.

But here’s the twist nobody saw coming: you can’t actually use these models yet. At least, not unless you’re one of roughly 20 pre-approved organizations that got the green light during a government-requested review period.

Why this launch is so unusual

Let’s start with the accessibility issue, because it’s the part that’s gotten most people talking. OpenAI announced Friday that GPT-5.6 Sol, Terra, and Luna are launching into a limited preview coordinated with the U.S. government. The company said it’s complying with the U.S. government’s request to initially limit the rollout to a small group of trusted partners.

OpenAI’s official line is that it “believes in broad access” and is working to make the models generally available in the coming weeks. During this preview period, the models are available through the API and Codex to select trusted partners and organizations.

This is the first time with a major model launch that you can’t use it on day one because the government asked OpenAI to hold the gate. The company explicitly stated they don’t believe this kind of government access process should become the long-term default, arguing it keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them.

Meet the three models: what makes each different

The GPT-5.6 family ships as three distinct models, named according to their capability tiers and inspired by the sun, earth, and moon. Here’s what each one does:

GPT-5.6 Sol is the flagship model, built for advanced reasoning, software development, and scientific applications. It’s OpenAI’s strongest model yet, setting a new state of the art on Terminal-Bench 2.1, which tests command-line workflows requiring planning, iteration, and tool coordination. On that benchmark, Sol Ultra scored 92%, while regular Sol hit 89%.

For coding workflows specifically, Sol shows broad improvements that make it the new performance leader. In biology workflows, it achieves stronger results than GPT-5.5 on GeneBench v1 while using fewer tokens. And for cybersecurity, it’s OpenAI’s most capable model yet, shifting the performance-efficiency frontier for long-horizon security tasks, including vulnerability research.

GPT-5.6 Terra is designed for enterprise and developer workloads, offering performance comparable to GPT-5.5 at roughly half the cost. It has competitive performance to GPT-5.5 while being 2x cheaper. Terra scored 84% on Terminal-Bench 2.1.

GPT-5.6 Luna is the fastest and most affordable model in the family, targeting cost-sensitive applications while retaining core capabilities. It brings strong capability at OpenAI’s lowest cost. Luna scored 82% on Terminal-Bench 2.1.

How much will this cost?

OpenAI introduced a new naming system with GPT-5.6, where the number identifies a model’s generation, while Sol, Terra, and Luna identify durable capability tiers that can advance on their own cadence.

The pricing per 1 million tokens is:

GPT-5.6 also introduces more predictable prompt caching, including support for explicit cache breakpoints and a 30-minute minimum cache life. For GPT-5.6 and later models, cache writes are billed at 1.25x the model’s uncached input rate, while cache reads continue to receive the 90% cached-input discount.

Safety first: the most robust stack yet

Sol launches with OpenAI’s most robust safety stack to date. The company strengthened protections for higher-risk activity, sensitive cyber requests, and repeated misuse, spending multiple weeks finding weaknesses, pressure-testing their system, and hardening it against real-world attacks.

They dedicated over 700,000 A100-equivalent GPU hours to automated red-teaming aimed at finding universal jailbreaks. This included model-level safeguards, real-time cyber and biology misuse classifiers, account-level review, and continued testing.

GPT-5.6 Sol does not cross the Cyber Critical threshold under OpenAI’s Preparedness Framework. In evaluations involving Chromium and Firefox, it identified bugs and exploitation primitives but did not autonomously produce a functional full-chain exploit under the conditions tested.

What’s next: when can regular people use this?

The company plans to make GPT-5.6 Sol, Terra, and Luna broadly available to people using ChatGPT, Codex, and the API soon. They’re excited to continue learning from this preview period and bring the models to more people soon.

OpenAI is also launching GPT-5.6 Sol on Cerebras at up to 750 tokens per second in July, bringing frontier intelligence to customers at unprecedented speed. Access will initially be limited to select customers as they expand capacity.