Updated on August 15, 2026. The prices on this page were redone with the current price list: OpenAI cut Luna by 80% and Terra by 20% on July 30, 2026. GPT-5.6 also left its restricted preview and has been open since July 9. If you read this post in June, the math has changed.
OpenAI just killed the idea of "the ChatGPT model." Now there are three.
On June 26, 2026, GPT-5.6 shipped, and it didn't arrive as a single model. It arrived as a family: Sol, Terra and Luna. Same generation, three capability tiers, three prices, three use cases. The developer's question stopped being "is it worth switching to 5.6?" and became "which of the three do I put in production?"
This post answers that: what the difference between GPT-5.6 Sol, Terra and Luna is, how much each one costs at today's prices, and when to use each one without burning money for nothing. No hype, just the engineering of the choice.
TL;DR
- What it is: GPT-5.6, OpenAI's current generation, split into three models by capability tier.
- The three: Sol (top), Terra (balanced), Luna (cheap and fast).
- Price per 1M tokens (input/output), July 30, 2026 price list: Sol $5/$30 · Terra $2/$12 · Luna $0.20/$1.20.
- The gap between the extremes: Sol costs exactly 25x Luna, on both input and output. In June it was 5x.
- Access: open. It left the restricted preview on July 9, 2026, and it's in ChatGPT, Codex and the API.
- If you budgeted with the old price: the redone math is in GPT-5.6 Luna's price dropped 80%.
Want to see Sol Ultra in action, with an independent benchmark? The post GPT-5.6 in Codex: Sol Ultra and the Terra gotcha has the first tests from outside OpenAI (CodeRabbit) comparing Sol and Terra on real coding tasks, and shows why Terra ends up more expensive than it looks.
The naming change (and why it matters to you)
Pay attention to this, because it changes how you pick a model from here on.
Before, the name was a single ruler: 4, 4o, 5, 5.5. The bigger the number, the better (and more expensive) the model. Simple and dumb.
With GPT-5.6, OpenAI separated the two things. The number (5.6) tells you the generation. The name (Sol, Terra or Luna) tells you the capability tier. And each tier evolves on its own cadence: a next-generation Sol could arrive before a new Terra, for example.
In practice, this forces you to think like an engineer, not like a consumer. You don't grab "the best model." You pick the right tier for each task: intelligence, speed or cost. It's the same logic as sizing infrastructure. You don't run everything on the most expensive machine just because it exists.
And this is where the conversation stops being about OpenAI. A tier router, evals on your own domain, cost measured per task: that isn't a vendor choice, it's architecture, and it's exactly the kind of system we build live, with code running, in the Clã Beer and Code. Tools and price lists change every three weeks; the decision criteria don't.
Sol: the top end, for when mistakes are expensive
Sol is the strongest model OpenAI has ever released. Its focus is heavy agentic work, with claimed gains in coding, biology and cybersecurity.
Two new features make Sol live up to its name:
maxreasoning effort: gives the model the maximum amount of time to reason deeply before answering. It's the "think more, cost more" setting.ultramode: goes beyond a single agent. It orchestrates subagents to break a complex problem into parts. This is OpenAI putting multi-agent architecture inside the model itself. If you use Codex, the first real tests are in GPT-5.6 Sol Ultra in Codex, including the side effect of it burning more tokens per task.
The number that got attention at launch: Sol Ultra hit 91.91% on Terminal-Bench 2.1, a record on the command-line automation benchmark. An honest caveat applies: it's OpenAI's own number. Treat it as a launch scenario, not as settled truth.
Price: $5 input / $30 output per 1M tokens. Unchanged since launch.
Use it when: the task is long, agentic and hard: a big refactor, terminal automation that can't fail, security analysis, a pipeline that makes decisions. Where getting it right is worth more than the bill at the end of the month. And remember that "getting it right is worth more" is no excuse for not measuring: we gave Sol a real business and it lied, spammed the users and burned through the cash.
Terra: the balanced one that probably becomes your default
Terra is the "balanced model for everyday work." And here's the point that matters most for production: it delivers performance competitive with GPT-5.5 at a fraction of the price.
Think about what that means. If you run 5.5 today, Terra is practically the same quality with a much smaller bill. It's not the tier that wins the benchmark headline. It's the one that wins on the invoice.
Price: $2 input / $12 output per 1M tokens (down 20% on July 30; it was $2.50/$15).
Use it when: it's your day-to-day production flow: text generation, medium-complexity code, user-facing chat, RAG. The place where you want high, consistent quality without paying flagship prices on every call.
A caveat the July cut created: Terra became the hardest tier to justify. It costs 10x Luna and is still well below Sol on hard tasks. It used to be the obvious middle ground; today, in a good share of cases, the honest decision is Luna or Sol, and Terra only comes in when the eval shows Luna can't handle it but Sol is overkill. Measure before assuming it's your default.
Luna: cheap and fast, for scale
Luna is the "fast and affordable" one. Strong capability at the lowest cost in the family. It's the tier you use when volume is high and depth of reasoning isn't the bottleneck.
Price: $0.20 input / $1.20 output per 1M tokens. Down 80% on July 30; it was $1/$6.
That cut is the single most important fact about the whole family. Twenty-five times cheaper than Sol, on both ends. In June the difference was 5x, and you could treat it as a spreadsheet detail. Twenty-five times is not a detail: it's an architecture decision that has to be defended in review, with numbers in hand.
Use it when: classification, structured extraction, moderation, routing, short chat replies, anything high-volume where latency and cost weigh more than the last drop of intelligence. Luna is the model that carries a feature firing off thousands of calls without blowing the budget.
An architecture pattern you can already sketch: Luna on the front line filtering and classifying, escalating to Sol only the cases that really need thinking. Routing by tier is how you avoid paying a lot for a cheap task.
Do not just follow the news — master it. Hands-on AI Engineering, live, every week, in the largest community in Brazil.
Join the ClãSummary: which GPT-5.6 model to choose
Price list in effect since July 30, 2026:
| Tier | Price (in/out per 1M) | vs Luna | What it's for | Use it when |
|---|---|---|---|---|
| Sol | $5 / $30 | 25x | Agentic flagship (coding, bio, cyber) + ultra mode | Hard, long task where accuracy > cost |
| Terra | $2 / $12 | 10x | Quality close to GPT-5.5, smaller bill | Middle ground, when the eval rules out Luna and Sol is overkill |
| Luna | $0.20 / $1.20 | — | Fast and cheap, for scale | High volume, routine work, cost/latency rule |
On a workload of 50M input tokens and 10M output tokens per month, that comes to $22 on Luna, $220 on Terra and $550 on Sol. Same task, three different orders of bill.
The detail that really changes the math: caching
List price is not the real cost. Anyone running LLMs in production knows that what defines the invoice is how much context you resend on every call.
GPT-5.6 brought more predictable prompt caching: you can set explicit cache breakpoints, and the cache has a minimum lifetime of 30 minutes. Cache writes are billed at 1.25x normal input, but reads keep the 90% discount.
Translated to product terms: if you have a large system prompt, fixed context or few-shot examples that repeat, you can lock that into the cache and pay almost nothing on the reads that follow. In an agent that makes dozens of calls with the same base context, this changes the economics more than choosing Luna over Terra. Well-designed caching is real cost optimization, not a nice-to-have.
Availability: what changed since launch
When GPT-5.6 came out on June 26, it shipped restricted: a preview open to about 20 organizations via the API and Codex, with the list shared with the United States government, a consequence of the executive order on evaluating frontier models. OpenAI itself complained publicly: "we don't believe this kind of government access process should become the long-term norm."
That's over. GPT-5.6 left preview on July 9, 2026, and all three tiers are available in ChatGPT, Codex and the API, including in Brazil. Three weeks later came the price cut.
What still stands as a caveat is something else: independent benchmarks are still scarce. Most of the numbers going around are OpenAI's own. Before migrating production because of a price list, put together a set of 30 to 50 real cases from your domain and run them on the tiers you're considering. A headline is not an eval.
The choice is engineering, not marketing
Sol, Terra and Luna in one sentence each:
- Sol thinks deep when the task is hard, and charges 25x for it.
- Terra is the middle ground that shrank after the July cut.
- Luna carries cheap scale, and got cheap enough to change the architecture.
The point isn't which one is "the best." It's that OpenAI gave you three tools and the responsibility of choosing. Whoever treats this as an architecture decision (routing by tier, using caching, measuring cost per task) pays less and ships better. Whoever throws everything at the most expensive model because "it's the top one" is outsourcing reasoning that should be theirs.
And the lesson from July is the one that sticks: the price list changed in three weeks, and it will change again. The AI bill can also get more expensive. Whoever built a provider abstraction switches tiers with one line in .env. Whoever hardcoded the model name in fifteen places in the codebase spends the afternoon refactoring.
FAQ
What's the difference between GPT-5.6 Sol, Terra and Luna?
They're three tiers of the same generation. Sol is the top, for hard agentic tasks, with max and ultra modes. Terra is the middle one. Luna is the fast, cheap one, for high volume. Same generation, different capabilities and prices: Sol costs 25x Luna.
Which one is the cheapest? Luna, at $0.20 input / $1.20 output per 1M tokens. That's 25 times cheaper than Sol on both ends, since the July 30, 2026 cut.
Is Luna good for coding? For simple, high-volume code, yes. For a heavy refactor or a long agentic task, the right tier is Sol. Luna was positioned for speed and cost, not for depth of reasoning.
Terra or Luna: which should I choose? It depends on the eval, not on the price. Terra costs 10x Luna. If your real cases pass on Luna, the difference in the bill is too big to ignore. If they don't, the next step up that's usually worth it is Sol, not Terra. That's what the July cut changed.
Can I use it in Brazil yet? Yes. Restricted access ended on July 9, 2026. All three tiers are in ChatGPT, Codex and the API.
What's the difference between max and ultra?
max is reasoning effort: it gives Sol more time to think deeply. ultra goes further: it uses subagents to break complex problems into parts. max is thinking more; ultra is coordinating several lines of reasoning.
{AI Engineer} — apaixonado por Laravel, arquitetura de software e construir produtos com impacto. Compartilho aqui tutoriais, descobertas e reflexões sobre o dia a dia de engenharia.
There is no shortage of content. What is missing is someone to untangle it: what matters now is how to implement it the right way. In the Clã you get that live, every week, with people who have already filtered out the noise.
Join the Clã