Quick answer: On July 9, 2026, OpenAI made GPT-5.6 generally available as three models. Sol is the top tier for hard reasoning and long agent tasks. Terra is the balanced everyday model, about as capable as GPT-5.5 at half the price. Luna is the fast, cheap one for high volume. The number 5.6 is the generation. Sol, Terra and Luna are capability tiers that can each improve on their own schedule. For most businesses, Terra is the sensible default, and you reach for Sol or Luna only where the work actually calls for it.
If you searched for GPT-5.6 this week, you probably ran into three unfamiliar names and a pricing table and left more confused than when you started. Here is the plain version. OpenAI stopped shipping one model per version and started shipping a small family. The release went live on July 9 after a short delay tied to a US government security review, and it is now available in ChatGPT, in Codex, and through the API.
The important shift is not the raw benchmark scores. It is that you now have to make a choice you did not have to make before: which tier fits which job. Get that choice right and you spend less while getting better answers. Get it wrong and you either overpay for simple tasks or send hard tasks to a model that is too small. This post walks through what each tier is for, what it costs, and how to decide.
What OpenAI actually released
GPT-5.6 is one generation of model offered in three sizes. OpenAI describes the naming this way: the number tells you the generation, and the name (Sol, Terra, Luna) tells you the capability tier. A tier can move forward on its own cadence, so a future Terra might get smarter without the whole family jumping a version number. That sounds like marketing, but it has a practical effect. You pick a tier for a job and mostly stay on it, instead of re-testing every model on every release.
Here is what each one is built for.
Sol, the flagship
Sol is the model for the hard stuff. Frontier reasoning, long-horizon agent work that runs across many steps, serious coding, and scientific or analytical problems where a wrong answer is expensive. If you are building an agent that plans and executes a multi-step task on its own, or you need the best available answer on a genuinely difficult question, this is the tier. OpenAI is also running Sol on Cerebras hardware at up to 750 tokens per second, which matters if you want frontier quality without a long wait.
Terra, the everyday workhorse
Terra is the one most teams will use most of the time. OpenAI positions it as roughly matching the older GPT-5.5 on quality while costing about half as much to run. For drafting, summarizing, answering customer questions, classifying tickets, and the general run of daily business tasks, Terra is fast enough, smart enough, and priced so you are not thinking about the bill every time someone runs a query.
Luna, the volume tier
Luna is the fastest and cheapest of the three. It is built for high volume and low latency: simple classification, short replies, tagging, routing, first-pass filtering, anything you run thousands of times a day where each individual call is easy. You would not hand Luna your hardest analysis, but for the long tail of small, repetitive calls it keeps costs down and responses quick.
What it costs
Pricing is per million tokens, split into input (what you send) and output (what the model writes back). A token is roughly three quarters of a word, so a million tokens is a lot of text, but it adds up fast at scale.
- Sol: $5 per million input tokens, $30 per million output tokens.
- Terra: $2.50 per million input tokens, $15 per million output tokens.
- Luna: $1 per million input tokens, $6 per million output tokens.
The gap is the whole point. Sol costs five times what Luna does on input and five times on output. Sending a task Luna could handle to Sol is not a small waste, it is a five times waste, repeated on every call. That is exactly why the tiering matters for anyone running these models at real volume.
There is also a quieter change that saves money if you use it: prompt caching got better. GPT-5.6 lets you set explicit cache breakpoints and keeps a cached prompt alive for at least 30 minutes. Reading from the cache still gets the 90 percent discount on input, while writing to it costs 1.25 times the normal input rate. In plain terms, if you send the same long instructions or context over and over, caching means you pay full price to store it once and a fraction of the price on every reuse for the next half hour.
Which tier should your business use
Do not pick one tier for everything. The point of a family is that you route each job to the cheapest tier that can do it well. A simple way to decide:
- Start on Terra. For most business tasks it is the right balance of quality and cost. Make it your default and only move off it for a reason.
- Move a task up to Sol when Terra visibly struggles: multi-step agent workflows, hard analysis, code that has to be right, or anything where a wrong answer costs real money. Test it, confirm the better answer is worth the higher price, then keep it there.
- Move a task down to Luna when it is simple and you run it constantly: tagging emails, routing tickets, short standard replies, filtering. The savings are largest exactly where the volume is highest.
You do not have to guess in the abstract. Pick one real workflow, run it on two tiers, and compare the output and the cost side by side. The right answer is usually obvious within an afternoon of testing, and it is different for different tasks. That per-task routing is where the savings live. We wrote more about controlling this kind of spend in our guide to cutting enterprise AI costs in 2026.
The other thing that shipped: ChatGPT Work
Alongside the models, OpenAI released a work-focused agent inside ChatGPT and tightened the controls for business accounts. Admins on Business, Enterprise, and Edu plans can set policies for which models, tools, and connectors staff are allowed to use, and cap usage. If you are rolling AI out across a team rather than to one power user, those controls are the part to pay attention to. They are how you keep a company deployment predictable instead of everyone using whatever they like on whatever budget.
If you are comparing options rather than committing to one vendor, it is worth reading these models next to the alternatives. We covered another recent release in our guide to Claude Fable 5 for business, and the sensible move for most companies is not loyalty to one lab but picking the right model for each job.
Should you switch right now?
If you already build on OpenAI, moving from GPT-5.5 to Terra is a straightforward upgrade to test, since Terra aims to match the older model at a lower price. There is a real saving on the table just from that swap, once you confirm quality holds on your own tasks. If you are on a competing model and it works, there is no urgency. A new release is a reason to run a quick comparison, not a reason to rip out something that already does the job. The steady approach beats chasing every launch: test on your actual workload, keep what wins, and route by task rather than by hype.
Frequently asked questions
What is the difference between Sol, Terra and Luna?
They are three sizes of the same GPT-5.6 generation. Sol is the most capable and most expensive, built for hard reasoning and long agent tasks. Terra is the balanced everyday model at about half the cost. Luna is the fastest and cheapest, made for high volume and simple tasks.
When did GPT-5.6 come out?
OpenAI previewed it on June 26, 2026 and made all three tiers generally available on July 9, 2026, after a short delay linked to a US government security review. It is available in ChatGPT, in Codex, and through the API.
How much does GPT-5.6 cost?
Per million tokens: Sol is $5 input and $30 output, Terra is $2.50 input and $15 output, and Luna is $1 input and $6 output. Prompt caching can lower the effective cost further if you reuse the same context often.
Which tier is best for a small business?
Start with Terra for most work. Move specific tasks up to Sol only when they are genuinely hard, and down to Luna when they are simple and run at high volume. Routing by task, rather than picking one tier for everything, is how you keep quality up and cost down.
Do I have to switch from GPT-5.5?
No, but Terra is worth testing, since it aims to match GPT-5.5 quality at a lower price. Run it on your own tasks first and switch only if the quality holds and the saving is real for you.
What is prompt caching and why does it matter?
If you send the same long instructions or context repeatedly, caching stores it so you do not pay full price every time. GPT-5.6 keeps a cached prompt for at least 30 minutes and gives a 90 percent discount on cached input, which cuts the bill for repetitive workloads.
Need help choosing and setting this up?
Buinsoft is a Prague-based AI and software consultancy. We help companies test these models on their real work, route each task to the tier that fits, and set up the access and cost controls before a rollout, not after the first surprise bill.
Write to us at info@buinsoft.com or use the contact form, and we will help you pick the right model for each job.
Details current as of July 10, 2026. Model pricing and availability can change, so check OpenAI's current documentation before you commit.
