ChatGPT-5.6 Luna, Terra or Sol? I Used Them for Months

01 / ARTIFICIAL INTELLIGENCE

GPT-5.6 Luna, Terra or Sol: What Is the Difference?

I went into GPT-5.6 expecting Sol to win by default. It does win some jobs. It also chews through a five-hour work window surprisingly quickly, and it still asks for another round when the brief is vague. After a few months, my answer became less exciting and more useful: I reach for Luna Max first.

That does not make the three models interchangeable. Luna, Terra and Sol have different prices, different amounts of reasoning effort and, judging from the Reddit comparison, very different ways of spending time. The model name alone is not enough; max, xhigh and high change the deal quite a bit.

That is the whole article in miniature. The rest is the receipt: what the Reddit chart actually shows, where Terra sits, why Luna Max works for me and when I would still pay for Sol.

02 / ARTIFICIAL INTELLIGENCE

The Reddit Cost Chart, Read Properly

The chart comes from a Reddit post comparing Pass@1, average task cost, output tokens and steps. Pass@1 is the percentage of tasks passed on the first attempt in that evaluation. It is not a guarantee that Sol will nail your next repository or that Luna will mess up your next email.

The 15 rows shown in the chart are listed below. A screenshot is handy for the overall shape, but a table makes the numbers much easier to search and compare. One footnote matters: this is a community comparison, not an official OpenAI price table and not a universal test.

The first thing that jumps out is Luna max: 67% at $0.61. Sol max is 73% at $8.39. So the chart gives Sol a six-point lead and roughly fourteen times the average cost. That is a big gap to pay for if the task is easy to review. It is a much smaller gap if the task can quietly break a production system.

The second thing is less flattering for simple rankings. Luna max uses 102 steps, while Sol max uses 61. Cheap does not mean quick, and a high score does not mean short. The model can arrive at the right answer by taking the long way round. Sometimes that is perfectly fine. Sometimes you have already gone for coffee twice.

03 / ARTIFICIAL INTELLIGENCE

What Each Model Is Good At

On the current GPT-5.6 page, OpenAI lists Sol at $5 per million input tokens and $30 per million output tokens, Terra at $2.50 and $15, and Luna at $1 and $6. Those numbers matter when you run the API. If you are using a work product with a rolling five-hour allowance, the effect is felt more as quota and waiting time than as a neat little invoice.

Here is how I read the three names in day-to-day work. Luna is the model I want when a human can spot the remaining error. Terra is useful once the job has a shape. Sol earns its keep when the hard part is deciding what to do, not merely doing more of it.

One small trap: people say they tested “Luna” or “Sol” as if each name describes one fixed experience. It does not. Luna low and Luna max are miles apart in the chart. The setting belongs in the comparison, otherwise you are comparing a car in first gear with the same car on the motorway and calling the result surprising.

04 / ARTIFICIAL INTELLIGENCE

The Chart Cost Is Not Your Account Limit

This is where a lot of model comparisons quietly go off the rails. A $0.61 Luna Max row does not mean your account will deduct sixty-one cents. An $8.39 Sol Max row does not arrive as an API bill either. Depending on the product, you may be dealing with messages, reasoning usage, a rolling window, plan-specific routing or several of those at once.

You still feel the difference. Heavy Sol runs leave less room for retries and other work. That is why I keep coming back to the five-hour limit: it is not an abstract pricing detail when you are in the middle of a workday and the model has just spent a large chunk of your available time thinking through a task that still needs a correction.

I think of it as three clocks. Money, model effort and available working time. They do not always tick together, which is why the cheapest row on a screenshot can still feel slow, while a premium run can feel expensive before you have seen a single dollar figure.

05 / ARTIFICIAL INTELLIGENCE

Why Luna Max Is the Interesting Option

The chart puts Luna Max at 67% Pass@1, $0.61, 73k output tokens and 102 steps. That is a strange combination. It uses more steps than Sol Max, but costs dramatically less. It is not elegant; it is useful.

Compare the nearby rows and the picture gets clearer. Luna xhigh is 57% at $0.31. Luna max is 67% at $0.61. In this test, an extra thirty cents buys ten points. Terra max reaches 70% at $3.96, while Sol high reaches 69% at $3.47. There is no smooth staircase where every extra dollar buys the same improvement.

So yes, Luna Max can wander. That is the trade. If the job is a draft, a testable patch or a table a person will inspect, I can live with the scenic route. If it is changing live records or making a decision nobody will review, I would rather spend more and add a human checkpoint anyway.

06 / ARTIFICIAL INTELLIGENCE

I Used Luna Max for Months

This is the part that matters most to me, because it came from work I was actually trying to finish. Luna Max does not always understand the brief the first time. I sometimes need to restate the tone, point at the exact section that went wrong or make the requirement painfully explicit. Two or three rounds is normal for demanding work.

The surprise was that Sol did not remove that relationship. It produced stronger first attempts on some tasks, especially when the problem was genuinely difficult, but I still had to steer it. And while I was doing that, my five-hour allowance disappeared much faster. That is a poor bargain for work where the final result still needs my judgement.

“Workability” is my own word here. I mean: can I keep moving without babysitting every sentence, and do I still have enough usage left for the rest of the day? Luna Max wins that test for me. Someone working on a difficult codebase may reasonably land somewhere else. Fair enough—different job, different bill.

07 / ARTIFICIAL INTELLIGENCE

When Sol Is Actually Worth It

The Reddit author’s routing idea is sensible: start with the lighter Luna settings, use Luna Max when the work gets sticky, then move through Sol high, Sol xhigh and Sol Max as the consequences get bigger. I would not open Sol Max just because the prompt looks important. A vague prompt can make a powerful model confidently spend time on the wrong job.

Sol high is probably the most interesting premium middle. The chart gives it 69% Pass@1, $3.47, 28k output tokens and 37 steps. Sol xhigh moves to 71% at $4.70. Sol Max reaches 73% at $8.39. That is a sensible ladder; you do not need to jump straight to the most expensive rung.

There is one catch, and it is not very dramatic: give Sol a proper brief. Files, limits, acceptance criteria, tests. Otherwise you are paying a specialist to improvise. Specialists can improvise, of course. That does not mean you will like the result.

08 / ARTIFICIAL INTELLIGENCE

Where Terra Fits

Terra Max scores 70% at $3.96 in the chart, with 72k output tokens and 76 steps. Terra xhigh is 60% at $1.70 and 43 steps. On those numbers, Sol high looks like a better middle deal for broad reasoning. Terra Max spends more than Sol high for one extra point in this particular evaluation.

The comments tell a different story for some users. One says Terra xhigh is faster than Sol. Another uses Terra for implementation and Sol for planning. Someone else prefers Luna xhigh because Sol tends to overthink straightforward tasks. I would not smooth those comments into one verdict; they are describing different kinds of work.

So I see Terra as the practical middle, not a consolation prize. It is the model I would try when the job has been planned but is too wide for a lightweight run. If the plan is still foggy, I would spend the budget on thinking first.

09 / ARTIFICIAL INTELLIGENCE

How I Choose a Model Now

1. Small and obviousLuna high or xhigh. The work is easy to inspect, so I do not need the expensive route.
2. Normal daily workLuna Max. This is where my own balance has held up.
3. Defined multi-file workTerra high or xhigh. I want more reach, but I already know the plan.
4. Unclear or high-risk workSol high. The hard part is judgement, not typing speed.
5. A stubborn or costly failureSol xhigh or Max. The premium is easier to justify once I know what the cheaper route could not solve.

This will not be everyone’s routing order. It is mine today. I would rather have a rule that can change after five real tests than a favourite model I defend because I saw one impressive screenshot.

10 / ARTIFICIAL INTELLIGENCE

If Luna Needs Two More Tries, Is It Still Cheaper?

If Luna gives me a workable draft and I only need to adjust the audience, one paragraph or a test case, two extra rounds are not a disaster. I can see what changed. The task remains cheap to check. That is where Luna earns its keep.

If every round is a full rerun, the model keeps bringing back the same defect or the output touches money, customers or production data, I would move up sooner. Saving on the model and spending the afternoon debugging its leftovers is not really saving. It just hides the invoice in your calendar.

I would measure cost per finished task, not cost per first answer. Include the review, the retries and the time spent explaining the same thing again. Your result may match the Reddit chart. It may not. Either outcome is useful because it belongs to your work.

11 / ARTIFICIAL INTELLIGENCE

Why Reddit Users Disagree

One commenter says Sol completed a hard project in one shot where several Luna sessions failed. Another says Sol overthinks and makes simple work harder. Someone prefers Terra xhigh for speed. Another uses Sol for planning and Terra for implementation. Those experiences can all be honest at the same time.

A writer cares about tone and whether the model obeys a small instruction. A developer cares about tests and whether the change breaks something three files away. A researcher may care about source tracking. Same model name, different pain point.

That last point is worth remembering. A chart is a snapshot. Keep it for the pattern, check the current OpenAI page for the product details and run a few of your own tasks before changing a workflow. Otherwise you are asking an old screenshot to explain a living product. It will not answer.

12 / ARTIFICIAL INTELLIGENCE

My Pick: Luna Max

So, which one should you pick? I would begin with Luna Max. Give it a proper brief, examples and a way to check the result. If it needs two or three alterations, that is not automatically a failure. Watch what those alterations cost you.

Move to Terra when the job is broader but already planned. Move to Sol when the work is genuinely ambiguous, difficult or expensive to get wrong. Sol is useful. I just do not think it should be running every ordinary task in the background like a sports car being used to buy groceries.

The test I would run

Pick five real tasks from one working week. Use the same files and prompt where possible. Record the first-pass result, number of revisions, waiting time, usage and whether you would actually send or ship the final output. That small log is more useful than arguing over a model leaderboard.

There is no prize for choosing the most expensive model. There is no prize for choosing the cheapest one either. Pay for the extra reasoning when you can point to what it bought you. Otherwise, keep the lighter model moving.