GPT-5.6 Explained (June 2026). Three Tiers.

OpenAI previewed GPT-5.6 on June 26, and the most important fact about it is one you cannot act on. The best version, Sol, is locked to roughly 20 approved companies by a US government order. The model everyone is talking about is one almost no one can use.

Look past the gate and the real news is the shape. GPT-5.6 is not one model, it is a three-tier family built for agentic, tool-heavy work, with a new mode that spins up subagents inside the model itself. Greg Brockman, OpenAI's president, summed it up in three words. "A good model." That is the understatement. Here is what actually changed, what it costs, and the part you need to know before you plan around it.

It Is Three Models, Named for the Sky

GPT-5.6 is not one model. It is a family of three: Sol, Terra, Luna. Sun, Earth, Moon. The naming is not decoration. It tells you exactly where to point each tier, and what each one costs.

The spread is deliberate. Sol is the frontier, the only tier with the new max and ultra reasoning modes, and the only one that is gated. Terra is the everyday default that quietly powers most Codex sessions. Luna is the volume play. You will almost certainly build on Terra and Luna, which is why the gate on Sol matters less than the headlines suggest. More on the gate below.

The Real Headline: Subagents in the Model

Sol ships with two new reasoning effort settings that sit above the old high. The first, max, gives the model the most time it has ever had to reason through a single problem before answering. That is the expected move, more thinking for harder questions.

The second is the one that matters. Ultra goes beyond a single agent. Instead of one model reasoning longer, ultra spins up subagents and parallelizes the work across them, then pulls the results back together. Orchestration that teams used to build by hand, with frameworks and glue code, is now a setting on the model itself.

Caching That Behaves

Less flashy, more useful day to day: prompt caching got predictable. GPT-5.6 adds explicit cache breakpoints, so you decide where the cached prefix ends, and a 30-minute minimum cache life, so a cached context does not silently evaporate between calls. Reads off the cache are 90% cheaper than fresh input.

If you run agents, this is the quiet win. Agent loops re-send the same instructions, tools, and context on every turn. With reliable caching and a real read discount, the standing context an agent carries stops being the expensive part of the bill. The model is finally priced the way agents actually use it.

The Benchmarks

OpenAI points to gains in coding, biology, and cybersecurity. The number worth holding onto is agentic: Sol sets a new state of the art on Terminal-Bench 2.1, the benchmark that tests command-line workflows requiring planning, iteration, and tool coordination. That is the right thing to measure for a model built to drive tools, not just answer questions.

The Catch: Sol Is Gated

Here is the part that changes how you should plan. GPT-5.6 launched as a limited preview, and the frontier Sol tier is not open. Per Dan Shipper of Every, a US government directive limited access to Sol to roughly 20 pre-approved companies on temporary national-security grounds. It mirrors the June export restrictions on Anthropic's most capable models. Terra and Luna are not gated, and general availability for all three is promised in the coming weeks.

What This Means for Your Company

Skip the temptation to chase the gated tier. The interesting move in GPT-5.6 is not Sol, it is the direction of travel: a model with subagents built in, a default tier good enough to run real work, and caching priced for agent loops. The frontier and everyday tiers are converging, and you will run whatever is good and available when you build.

So the model is not the decision. The decision is whether your functions are ready for one. Clean context, defined goals, connected data, a clear definition of done. Get that right for one function and the same playbook repeats across the next, whatever model is sitting underneath. That is the work, and it is the work nativefirst does on site.