GPT-5.6 Sol, Terra, and Luna go generally available

# The three-tier stack ships from the source

OpenAI made GPT-5.6 generally available on July 9. Three tiers — Sol, Terra, Luna — one 1.05M-token context window, three prices. It is the first frontier release I have seen where the pricing sheet reads like a product manager, not a research lab.

Sol is $5 in and $30 out per million tokens. Terra is $2.50 in and $15 out. Luna is $1 in and $6 out. Sol Ultra, the high-compute mode, sits on top of Sol as a switch, not a separate SKU. Sol Ultra scores 91.9 on OpenAI's composite. Sol scores 88.8. Terra scores 84.3, which puts it a hair above GPT-5.5 at 83.4 while costing roughly half. Luna scores 82.5.

If you have been reading these Briefs for a while, you know I have been repeating the same slide for two years. Your AI stack should have three tiers. A frontier model for the five percent of hard calls. A middle model for the eighty percent of routine calls. A heuristic or a narrow model for the fifteen percent that never should have been paying LLM prices. Most teams I audit are paying frontier for all three.

OpenAI has now built that stack for you. Which is helpful and also worth watching. Helpful because the choice architecture is now obvious to any engineer opening the pricing page. Worth watching because when the same lab defines the frontier, the middle, and the floor, the cost curve for the middle tier is the one you should track. Terra is where enterprise budgets will spend most of the next year.

The Sol-on-Cerebras deployment is the other half of the story, and the half I find more interesting. Cerebras is now serving Sol at up to 750 tokens per second on wafer-scale hardware. A comparable H100 cluster serves a frontier model at roughly 70 tokens per second. That is about ten times faster on inference. Independent analysts estimate the deployment uses seventy to one hundred wafers with roughly one model layer per wafer, which is the kind of lavish physical layout that only makes sense if you have decided your agent's wall-clock latency is now a first-class product constraint.

That is the part worth saying out loud. When you are running a five-step agent trajectory, the difference between 70 and 750 tokens per second is not a benchmark, it is whether the agent finishes while the user is still looking at the screen. The teams I advise who are shipping agentic surfaces have all had the same conversation this year: at frontier price and frontier latency, the agent product is a demo. The Cerebras deployment says OpenAI has decided to fix the latency side of that trade for the customers who will pay for it. Expect the pricing on that path to be non-public and steep.

A few honest cautions.

The 1.05M context window is real, but the useful working span is smaller than the marketing number. I have not run my own attention-degradation tests yet. I will report back. If you are building on top of the long context, run your own recall evals on the second half of the window before you commit.

The Sol Ultra scores are OpenAI's own composite. Artificial Analysis and the other independent leaderboards will land over the next two weeks. I would not price a contract on the Ultra number until at least two independent replications confirm it.

Terra at half the price of GPT-5.5 is the number that will move the most product roadmaps. If you have a production surface running on GPT-5.5, budget an afternoon this week to run your top-200 queries against Terra. Not the adversarial set. The boring cases. If Terra ties or wins, migrate. If it loses on a specific slice, keep GPT-5.5 for that slice and route the rest.

Two things I will be watching in the next month.

First, whether Luna is priced low enough to eat the middle of the small-model market. At $1 in and $6 out on a million-token context, it competes directly with the open-weight middleweights. If it does, the 'run your own 7B' argument gets narrower fast.

Second, whether Sol Ultra plus Cerebras is the first frontier tier where latency, not capability, is the differentiator. If it is, we are going to need a new column in every eval table. Tokens per second per dollar on a real trajectory is the number I would put on the roadmap slide.

The three-tier stack is not a new idea. Shipping it as a pricing sheet is. The burden now shifts to product teams to actually route the traffic instead of paying frontier for everything.

What are you migrating this week?