GPT-6 Sol vs Luna: which one to choose and for which jobs

GPT-6 Sol and GPT-6 Luna are the two models that OpenAI added to GPT-6 Astra on September 22, 2026. The choice between the two boils down to a question: does work have an answer that you can quickly control? If yes, like when you extract the fields of an invoice or summarize an email, start from Luna. If the task asks to reason on several steps, such as writing or reviewing code or building an analysis, start from Sol.

The comparison is unusual because on paper the two models are almost twin: same 1.050,000 token context window, same maximum output, same levels of reasoning, same tools. Change the price, Twenty times on each line of the price list, and changes what OpenAI promises on each. Verified data on September 23, 2026 on OpenAI API documentation: before budgeting a project always double-check the official page.

What GPT-6 Sol and Luna have in common, and what not

In the official cards of GPT-6 Sol and of GPT-6 Luna The differences lie in a few lines: price, stated level of reasoning and limits of demands.

Two details of the cards say how OpenAI imagines the use of the two models. Both have the same speed label, so speed is not the reason to choose Luna. And the limits of Luna are wider: already at Tier 2 it accepts twice the tokens per minute of Sol, at Tier 5 more than quadruple, consistent with a model designed for large volumes. Those who start from scratch find the overall picture, including ChatGPT plans, in What are GPT-6 Sol and Luna.

Twenty times less: where you see it on the price list

The ratio of the two prices is the same on each row of the official pricing: input, read and write cache, output, long context, and also Batch, Flex and Fast modes. For a reference, the predecessor GPT-5.6 Luna cost $0.20 and $1.20 and we had put it next to Claude Haiku 4.5 in this comparison.

Two rules apply to both and move the account more than the choice of the model. Above 272,000 input tokens the entire request is paid double the input and one and a half times the output, as indicated by the model cards: Sol goes to $4 and $15, Luna to $0.20 and $0.75. And the reasoning tokens, even if you don't see them, are paid as output tokens. Raising the effort on Luna costs less than on Sol, but it is not free.

How far is Luna from Sol in OpenAI tests

The numbers come from the September 22 announcement and it measured them OpenAI itself. In each line look at the level of effort: the same model, at different effort, gives different results.

The line that really answers the "Sol or Luna" question is one, DeepSWE v1.1: 68.8 versus 66.6, same test, same effort. Just over two points away with a price per token twenty times lower. It's a software development test, so it doesn't say anything direct on invoices or emails, but it suggests that there's less distance between the two than the list shows.

The other lines should be read more carefully. For Luna OpenAI it measures the jump over the GPT-5.6 models, not the distance from GPT-6 Sol. And the fact assessment uses conversations where users had reported an error, which OpenAI itself declares non-representative of normal use. The reason for this caution is in because vendor benchmarks are not comparable, the weight of effort in reasoning effort: what changes with regard to.

Curiosity: Luna knows more recent things than Sol

In the previous family the three variants of GPT-5.6 had the same limit date, February 16, 2026, as we had reported in GPT-5.6 Sol vs Terra. With GPT-6 dates separate, and contrary to what would be expected: knowledge of Luna arrives May 18, 2026, that of Sol stops at April 20, 2026. The cheapest model is the most up-to-date, and OpenAI does not explain the reason.

For an SME it weighs little: four more weeks are not enough to trust the model's memory on prices, rules or deadlines. For recent facts you pass the right documents or activate web search, available for both. The theme is deepened in knowledge cutoff: what it means in practice.

How much does the same job cost with Sol and Luna

Three examples with Standard prices, required below 272,000 tokens and no cache. Volumes and tokens are assumptions to show the order of magnitude, not measurements made on a company.

The first and third cases say the same thing: when the bill is below $40 a month, between the two models they dance a few tens of dollars, and weighs more the quality of the result. There you choose on that. The second case is the one where the list really decides: in a year they make about $200 with Luna and almost $4,000 with Sol, for a job that Luna is built to do.

Two fixes bring examples closer to the real account. Repeated fixed instructions at each call, cached, cost one tenth of the entry price, according to the Luna's card. The raised effort adds reasoning tokens to the exit. With Luna the two effects shift cents, with Sol tens of dollars.

The work of an SME, divided between Luna and Sol

The table starts from jobs, not models. The right column says which signal, measured on your cases, justifies a change.

A distributor of electrical equipment with 150 orders per day. Think of a company receiving orders by email, half in PDFs and half written in the message. Luna reads each order and fills in codes, quantities and delivery dates in the business management software. Orders with a code absent from the price list or an out-of-range quantity go into a separate queue, where Sol compares them with the customer’s historical records and proposes a correction to the approver. The first half of the workflow is shown in the demo by automatic data entry.

A software house of six people. This is where the report flips. Sol works in Codex on new features and revisions before merge, where an escaped bug costs more than a month's consumption. Luna makes the outline: summarizes the open issues, writes the release notes, generates test data. Details are in Codex with GPT-6 Sol and Luna.

When Luna is not enough, and how to notice it first

You need a brief test on your worst cases. Take thirty real cases per stream, with crooked scans and hastily written emails, and spin Luna with production instructions. Then raise the effort and repeat: only what remains wrong is matter for Sol.

There is a third road, often the cheapest: Luna makes the big one and passes to Sol only the uncertain cases. Here it works well because the two models share context, APIs and tools, so moving a step does not change the architecture: the schema is in Mixed Model Strategy. If even Sol is not enough, above is GPT-6 Astra, the best model of OpenAI, compared to the direct rival in GPT-6 Astra vs Fable 5.1.

Five mistakes that inflate the bill

1. Put Sol everywhere out of caution. On extraction and summaries you pay twenty times as much for a margin that in your cases may not be there.
2. Compare the two models to different efforts. Luna at none versus Sol at medium tests effort, not the models. Set the same level and increase it only for the model that loses.
3. Forget the reasoning tokens. They are paid as output even if they do not appear in the answer: measure them on your cases before the quote.
4. Exceed the 272,000 tokens without realizing it. One attachment too many and all the request passes to the price of the long context. With huge documents only sends the useful parts.
5. Try only on clean documents. The distance between the two models is seen on scans, broken tables and ambiguous requests. They are the ones to put in the test.

Sol or Luna: the rule to take away

Luna for what is repeated and controlled, Sol for what needs to be reasoned and then signed. With the same context, the same tools and a different price twenty times, the useful question is not which of the two is better, but how many of your steps can go down to Luna without losing quality. The answer comes from a test on your cases, not from the list.

If you also evaluate the Anthropic alternative released on the same day, the comparison is in GPT-6 Sol vs Claude Opus 5.5. And if you want to put Luna and Sol in the right place inside a stream that already exists, orders, tickets or code revisions, with Giallo Studio Let's start with your real papers and measure before we decide.