GPT-5.6 Is Here: Sol vs. Terra vs. Luna Explained
A plain guide to the three new OpenAI models, what they cost, and which one to pick.
OpenAI just changed ChatGPT again, but this update is easier to understand than the old model menu.
GPT-5.6 comes in three sizes: Sol, Terra, and Luna. One is built for the hardest work. One is the smart middle choice. One is made for speed and low cost.
The first guides about GPT-5.6 came out during a small preview in June. That is when this deep guide from Tosea was written. But the access news changed on July 9, 2026. OpenAI says GPT-5.6 is now rolling out across ChatGPT, Codex, and the API.
So here is the new version of the story, in plain English.
The fast answer
If you only remember one thing, remember this:
Use Sol for hard work. Pick it for deep research, big coding jobs, hard plans, and work where the answer needs more thought.
Use Terra for most daily work. It is the best place to start when you want strong work without paying Sol prices.
Use Luna for fast, simple work. Pick it for sorting, short drafts, quick checks, and jobs you need to run many times.
Most people do not need Sol for every chat. Start with Terra. Move up to Sol when the task is hard. Move down to Luna when speed and cost come first.
Sol is the heavy thinker
Sol is the top GPT-5.6 model. OpenAI calls it the flagship model.
This is the one to use when a task has many steps and a bad answer would waste time. Think about a large code fix, a deep research report, a full business plan, or a messy file set that needs to be read and checked.
Sol also has two new work modes. Max gives the model more time to think and test its answer. Ultra uses several agents at once, then pulls their work together. In Codex, OpenAI says Ultra is available to Plus plans and above. In ChatGPT Work, it is for Pro and Enterprise users.
Ultra can use more tokens, so it is not the right button for every tiny job. Save it for work that is worth the extra effort.
Terra is the smart default
Terra is the model I would tell most people to try first.
OpenAI says Terra has skill close to GPT-5.5, but its API price is half as much. That makes it a good fit for daily writing, planning, code help, research, and business tasks.
Terra is also the model that Free and Go users get in ChatGPT Work and Codex. Paid users can pick Sol, Terra, or Luna.
If you run a small business, Terra can handle a lot of the work you do each day. It can help turn notes into a plan, clean up copy, study a file, make a task list, or draft a first version. You can send the hardest parts to Sol after Terra gets the job started.
Luna is built for speed
Luna is the fastest and lowest-cost model in the group.
It is a good fit for short jobs that do not need deep thought. You can use it to tag leads, sort feedback, clean up a list, pull fields from text, or make many short drafts.
Luna is not the model I would pick for a hard choice or a long plan. But it can be a great worker inside an automation because small jobs can add up fast.
This is the part many people miss. The best model is not always the biggest one. The best model is the least costly model that can do the job well.
What the API costs
OpenAI lists these prices for one million tokens:
Sol: $5 for input and $30 for output
Terra: $2.50 for input and $15 for output
Luna: $1 for input and $6 for output
Input is what you send to the model. Output is what the model sends back.
For many small teams, Terra will be the main model. Luna can handle simple steps. Sol can step in when the work gets hard. This kind of model routing can cut costs without making every result worse.
OpenAI also says all three API models have a 1,050,000-token context window and can return up to 128,000 tokens. That is a lot of room for files and long jobs. Some early articles said the context window was 1.5 million tokens, but the current official model pages say 1.05 million.
A simple way to pick your model
Use this three-step check before you start:
How bad would a wrong answer be? Use Sol when the risk is high.
How hard is the work? Use Terra for most daily work and Luna for simple steps.
How many times will this run? Cost gets more important when a task runs all day.
You can also build a ladder. Start a job with Luna. Send the hard parts to Terra. Send only the hardest cases to Sol. That is often better than using the biggest model for every step.
The benchmark numbers need context
OpenAI reports that Sol scored 88.8% on Terminal-Bench 2.1, a test for command-line work. Sol with Ultra reached 91.9%. Terra scored 87.4%, and Luna scored 84.7%.
Those are OpenAI's own reported numbers. They are useful, but they do not prove which model will work best for your real job. Ultra also uses several agents, so its score is not a clean one-model test.
The best test is still your work. Give each model the same real task. Check the answer, time, and cost. Then pick the smallest model that passes.
What to do now
The rollout may take up to 24 hours, so you may not see all three models at once.
When they show up, do not test them with a random question. Pick one real job you already do. Run it through Luna, Terra, and Sol. Keep the prompt and files the same. Then compare the work.
My bet is that Terra will be enough for most daily jobs. Sol will earn its place on the hard ones. Luna will do more of the quiet work behind the scenes.
The names are new, but the rule is simple: match the model to the job.
Sources
If you test the new models, leave a comment and tell me which one won your real task.