GPT-6 Sol and Luna: Same job at half the price?

What's in this month's feature? Go on, guess. Hint: there's a sun in the name, and a moon too. On 22 September 2026 OpenAI announced the two little siblings of the GPT-6 family, Sol and Luna. They replace GPT-5.6 Sol and GPT-5.6 Luna and sit one step below GPT-6 Astra, the flagship that came out at the start of September. Let me say this right up front: this piece is built on OpenAI's pricing page, the model documentation and the launch coverage. Nobody ran these models for this magazine. I have no measurements, and I'm not going to invent the ones I don't have.

Let's not drag this out. On with the job.

Who's who?

The split is actually pretty simple. Sol is built for complex work, coding above all. Luna takes the quick jobs that arrive by the truckload: summarising, pulling data out of a text, short answers to short questions. So one is the master craftsman and the other is the apprentice. But this apprentice works all day without a break and hardly asks to be paid.

The two have the same spec sheet:

• Context window: about 1.05 million tokens. Of that, 922 thousand go to input and 128 thousand to output.
• Input and output: text and images go in, only text comes out.
• Knowledge cutoff: 20 April 2026 for Sol, 18 May 2026 for Luna. The little sibling is a month more up to date than big brother. Cute detail.
• Reasoning effort: both support the none setting, so you can tell them "don't think, just answer".

Think for a moment about what a million tokens means. Once upon a time a whole floppy carried a game. Now you can feed almost an entire medium-sized code repository to the model in a single request. Feeding it in is one thing, doing justice to what you fed in is another, of course (I'll get to that below).

Their API names are gpt-6-sol and gpt-6-luna. On the ChatGPT Work and Codex side they've been rolling out gradually to Plus, Pro, Business, Enterprise and Edu users. Free and Go users can also try Luna in the desktop app. According to TechCrunch, the switchover in the ChatGPT app and on the website happened bit by bit over launch day.

Now for the real business: money

These two models exist because of the price. OpenAI doesn't hide it either. The figures are per million tokens, on the Standard tier, for prompts up to 272 thousand tokens:

• gpt-6-sol: input $2.00, cached input $0.20, cache write $2.50, output $10.00.
• gpt-6-luna: input $0.10, cached input $0.01, cache write $0.125, output $0.50.

Now let's line them up against the old ones. GPT-5.6 Sol wanted $4 for input and $20 for output, so Sol's price has been cut exactly in half. GPT-5.6 Luna was $0.20 and $1.20. Luna's input has been halved, and its output has dropped below even half. Astra is still sitting at $10 and $50. So Sol's output is five times cheaper than Astra's. OpenAI pitches the cut as 50 percent off GPT-5.6's promotional prices, and credits better caching and more efficient inference.

I have to open a parenthesis here (if I don't, your bill will open your eyes for you). The price comes with footnotes:

• On prompts over 272 thousand tokens, input doubles and output goes up one and a half times.
• The Batch and Flex tiers are half the Standard price.
• Fast mode is double.
• Regional processing adds 10 percent.

So if you fill that million-token window to the brim, the "half price" fairy tale loses a lot of its sparkle. The window is big, but bigness has its own tariff. Finding that out without reading the docs would be a pretty expensive lesson. Luckily OpenAI spells it all out on the pricing page, credit where it's due.

So what did we gain, and what did we lose?

This is where it gets interesting. OpenAI's biggest claim isn't about speed or brains, it's about accuracy. By the company's own account, on an internal accuracy test built from answers users had flagged as wrong, Sol makes about half as many mistakes as GPT-5.6 Sol. OpenAI is selling this as "Astra-level reliability at a much lower price". Second-hand sites are passing around figures like 4.6 percent against 8.5 percent at the highest effort, and another one says 4.5 percent at xhigh. I couldn't verify those exact figures in a primary source, so you should read them as rumours too. The part that can be verified is this: the test is OpenAI's own, and nobody independent has repeated it.

On top of that, the company says the error rate has dropped on coding tasks and that Sol beats Anthropic's top models (Fable and Opus). As TechCrunch also points out, every OpenAI launch comes with a sentence like that. Let's file it under vendor claims and move on.

The place you really need to pay attention is somewhere else. OpenAI's own tables (again, as second-hand sites report them) show Sol falling behind the old Sol on the top scores. 68.8 percent against 72.7 percent on DeepSWE, 64.4 percent against 66.2 percent on OSWorld 2.0. For Luna, Artificial Analysis's independent rating is also coming through second-hand sources, and I couldn't check it at the source. According to that rating GPT-6 Luna (max) scored 41, two points below GPT-5.6 Luna. There are small drops on the coding tests too: 44 percent against 49 percent on SWE-Atlas-QnA, 64 percent against 66 percent on DeepSWE.

So here's the picture: the version number went up by a whole point, but the abilities didn't go up, they even slipped a point or two. The gain is on the invoice and in accuracy. A model with "6" on the label knowing less than "5.6" is a bit funny, isn't it? Actually, not that funny. These models weren't built to be smarter. They were built to be wrong less often and to eat less money. The marketing department picks version numbers, not the engineers. Read it knowing that and you won't be disappointed.

No serious review has come out of the big outlets yet. Most of the coverage out there just repeats OpenAI's numbers, and that's that. The models have only been out for a month, and independent measurements will need a while to settle.

Who for, and what for?

Here's the verdict I draw from the docs and the price:

Luna really is pretty tempting for anyone who chews through text by the ton. This is a model that reads a million tokens for 10 cents and serves cached input for 1 cent. For a job that classifies thousands of support tickets a day or pulls fields out of hundreds of documents, the sums change completely. Nobody notices a two-point drop in brains on that kind of work. What gets noticed is the bill at the end of the month.

Sol, on the other hand, is a middle road, but a well-balanced one. It falls a little behind the old Sol on the hardest coding tasks, and the smart move is still to keep Astra in a slot of its own. But if "giving wrong information" is your real headache, and OpenAI's claim holds up, you're getting a model that's wrong less often at half the price. For work that boils documents down, that gives customers information, where a made-up answer costs you dearly, the scales tip towards Sol.

There's exactly one point where I won't budge: before you move a job that's running in production to the new model, you MUST run a comparison on your own examples! With the top scores down, shrugging "it's the new version, so it must be better" is exactly the recipe for waking up one morning to find everything has crashed. Keep the old model's answers somewhere, put the new one's next to them. Then decide.

A few small oddities

Before I wrap up, let me be honest and note down a few things.

First, OpenAI's own announcement page was closed to us. I couldn't read its text directly, so I gathered the information from the announcement on the developer community, the pricing page, the model pages and TechCrunch. Second, one search result says the date 10 September shows up in OpenAI's research index. I don't know why. Every other source says 22 September, so that's the date I went with.

The third one is pretty entertaining: as I write these lines, OpenAI's developer docs already list a newer model called GPT-6.1 Sol. Its price is again $2 and $10, but cached input is $0.10, so half of even this one's. I couldn't verify when it came out. A model whose successor turns up before it's a month old... Back in the day a single version of a program went from hand to hand for years. Now a magazine article can outlive the model it's about.

Then there's the timing. According to TechCrunch, Sol and Luna were announced about 90 minutes after Anthropic released Claude Opus 5.5. I'm noting that purely as a fact. The who-was-chasing-whom debate I leave to the stands (they've got enough on their plate as it is).

If you've used one of these models in your own work, write and tell us what you saw. Letter, e-mail, comment, whichever fits your hand best. When the independent measurements settle and the 6.1 story gets clearer, I'll reopen this file. Promise.

Right then, see you next issue, take care!..