What is GPT-5.6? Sol, Terra and Luna explained simply

If you've seen the names Sol, Terra and Luna floating around and wondered what on earth OpenAI is up to, you're in exactly the right place. On 26 June 2026, OpenAI previewed GPT-5.6, its most capable family of models yet. The twist that got everyone talking isn't the technology at all: it's that, for the first time, a US lab has held its best model behind a government-approved access gate. Let's unpack all of it in plain English, and work out what it means for you.

Think of GPT-5.6 not as a single product but as a small family with three members, each pitched at a different job and budget. Sol is the flagship: the muscle you'd reach for on the hardest reasoning, coding, agentic and security tasks. Terra is the sensible all-rounder, delivering roughly the performance of the previous GPT-5.5 at about half the cost, which makes it the natural pick for everyday work. Luna is the budget tier, still genuinely capable but built to keep costs as low as possible when you're running a lot of requests. It's a bit like a car range with a performance model, a comfortable family model and an economical commuter: same badge, different priorities.

Pricing is measured per 1 million tokens (tokens are the little chunks of text a model reads and writes — very roughly, a token is about three-quarters of a word). There are two numbers for each model: what you pay for the text you send in, and what you pay for the text it generates out. Here's the whole range at a glance:

The pattern is friendly and predictable: each step down the ladder roughly halves the price. Most people and teams will happily live on Terra or Luna, reaching for Sol only when a task genuinely needs the extra horsepower — which is a nice change from feeling you have to pay flagship prices for everything.

Here's the part that surprised a lot of people. Normally a new OpenAI model shows up in ChatGPT within days. Not this time. During the preview, GPT-5.6 is available only through the OpenAI API and Codex — the developer-facing tools — and not in ChatGPT at all. More striking still, it went to just about 20 partners, and OpenAI didn't pick them alone: the participants were approved by the US government.

Why the caution? It traces back to a Trump executive order signed on 2 June 2026, which directs government agencies to build a framework for benchmarking and assessing powerful new AI models and to designate so-called "covered frontier models" — systems with advanced cyber capabilities — before they're released broadly. GPT-5.6 is the first flagship to go through that new gate, which is why the rollout looks so unusual.

On the safety question specifically, OpenAI says that under its Preparedness Framework it treats all three models as High capability in cybersecurity and biological/chemical areas, while none reaches High in AI self-improvement. On cyber in particular, it reckons Sol is better at helping people find and fix vulnerabilities than at reliably carrying out end-to-end attacks, and doesn't hit the "critical" threshold. In other words, powerful enough to warrant a careful look — which is exactly what the review is for.

The encouraging news is that this is a starting point, not a permanent lockout. OpenAI expects to expand access in the coming weeks, so wider availability — and, in time, a route into the tools you already use — is on the way. It's also worth knowing this isn't a one-off targeting a single company: the same federal machinery briefly pulled Anthropic's newest models in June, so the access gate is becoming a shared pattern across the industry rather than one lab's problem.

A couple of extras worth a mention if you like the shiny bits: there's a new "ultra" mode that splits a task across multiple sub-agents working in parallel, and OpenAI plans to run GPT-5.6 Sol on Cerebras hardware at up to 750 tokens per second in July, initially for select customers. Blisteringly fast, in other words — once the gate opens a little wider.