ISA After Hours Augsburg
An office at dusk seen from inside. Through a tall window above a city skyline, the sun sets in orange on one side of the sky while a pale crescent moon has already risen on the other. A closed laptop, a notebook and a pen sit on a wooden desk in the last light.

AI at work · New models

GPT-6 Sol and Luna: top AI, half price

· About 8 min read

On the same day Anthropic cut the price of its top model, OpenAI made its own move. It took what it learned building GPT-6 Astra, its most capable model, put it into two cheaper models called Sol and Luna, and cut their prices in half. The goal, in OpenAI's words: "More ways to bring frontier intelligence into the work you do every day."

ISA After Hours Augsburg · All posts

In this article
  1. What OpenAI released
  2. Half the price
  3. Three models, three jobs
  4. What they can do
  5. What got better
  6. Official sources

Short version

GPT-6 Sol and GPT-6 Luna are OpenAI's new everyday models, released on 22 September 2026. They were trained with methods similar to GPT-6 Astra's and cost half as much as the models they replace. Sol is the workhorse for demanding daily work, Luna is the cheap and fast one, and Astra stays the top model for the hardest jobs. Both new models make fewer factual mistakes and write clearer, shorter answers. They are in ChatGPT and Codex for paying plans from today, and free users get Luna in the desktop app.

What OpenAI released, and why

Three weeks ago OpenAI launched GPT-6 Astra, which it calls "the most intelligent and aligned model in the world." Astra is powerful and priced to match. OpenAI's own launch page explains why that is not enough: "While the most demanding and important projects still call for Astra's full depth, work happens at different scales, rhythms, and budgets."

So OpenAI trained two smaller models "with similar methods as GPT-6 Astra", bringing the advances behind Astra to "faster, more affordable models." In OpenAI's words, the new pair "help distribute the benefits of that intelligence by advancing the frontier on cost efficiency." Put simply, the best model's way of working now comes in cheaper sizes.

The timing is part of the story. Sol and Luna were announced on 22 September at 18:00 GMT, according to OpenAI's news feed, the same day Anthropic released Claude Opus 5.5 and cut its prices by 40%. Two of the biggest AI makers moved in the same direction within hours: top-level work, at a lower price.

What this means for you: the most useful AI news this month is not a new record. It is that the good models are getting cheap enough to use for everything.

Half the price: Sol and Luna in plain money

Prices are quoted per million tokens. A token is a piece of a word, roughly three quarters of an English word. "In" is what you send the model, "out" is what it writes back. OpenAI cut both, for both models, by half compared with the promotional prices of the GPT-5.6 versions they replace.

Per million tokensBefore (GPT-5.6)Now (GPT-6)Change
Sol, in / out $4 / $20 $2 / $10 50% cheaper
Luna, in / out $0.20 / $1.20 $0.10 / $0.50 50% cheaper on input, more on output

Old and new prices from OpenAI's launch page, new prices confirmed on OpenAI's developer pages for Sol and Luna, 23 September 2026.

OpenAI also made re-reading cheaper. On a long job an AI reads the same material again and again: your documents, the conversation so far. It keeps that material in a short-term memory called a cache, and re-reading from it costs far less than reading fresh. OpenAI says GPT-6 now finds more of what it has already read by default, with "discounts of 90% on cached input-token reads." For Sol, that is $0.20 per million tokens instead of $2. GitHub reports that these improvements cut the share of text that needs fresh processing by more than 50%, which made Copilot answer faster.

Why does this matter so much? OpenAI gives a telling figure from its own offices. As AI coding agents take on longer work, daily usage valued at API prices has exceeded $600 for the median OpenAI researcher, and $7,000 for the heaviest users. At that scale the price per token decides what people dare to ask for.

For comparison, Claude Opus 5.5, released the same day, costs $4 in and $20 out, so Sol is half its price per token. That does not tell you which is better value. Each company's launch compares itself with the other's previous models: OpenAI measures Sol against Claude Opus 5 and Fable 5.1, and Anthropic measures Opus 5.5 against GPT-5.6 Sol and GPT-6 Astra. A direct comparison of the two new models does not exist yet.

What this means for you: for most office work the cost of AI is now small change. The expensive part is choosing the right model for the job, which is the next section.

Three models, three jobs: Astra, Sol, Luna

OpenAI now sells one family in three sizes. The idea is the same as with any team: you do not send your most senior expert to answer every email, and you do not give a trainee the board presentation.

ModelPrice per million tokens, in / outWhat it is for
GPT-6 Astra $10 / $50 The hardest, most important work. OpenAI: "Choose it when you want the best results and an uncompromising experience."
GPT-6 Sol $2 / $10 Difficult everyday work, with "more room to iterate with higher usage limits and lower cost"
GPT-6 Luna $0.10 / $0.50 Fast, cheap, high-volume work, and the model free users get

Prices from OpenAI's developer pages for Astra, Sol and Luna, checked on 23 September 2026.

Where you find them depends on your plan. Sol and Luna are available in ChatGPT Work and Codex, OpenAI's coding agent, for Plus, Pro, Business, Enterprise and Edu users. Free and Go users can use Luna in the ChatGPT desktop app. OpenAI notes that the models are "not yet available in Chat", the ordinary conversation side of ChatGPT, and that they roll out gradually through the day, so they may appear a little later for you. Developers call them as gpt-6-sol and gpt-6-luna.

What this means for you: start with Sol for real work. Move up to Astra when the result really matters and Sol falls short, and use Luna for quick, repetitive jobs where speed and cost count more than depth.

What they can do

OpenAI tested the new models on the kind of work people do on computers. All results below are OpenAI's own measurements, and the competitor scores come from public reports, as OpenAI notes.

Office workflows. AutomationBench tests whether an AI agent can carry out a business process from start to finish, using 47 tools across sales, marketing, operations, support, finance and HR. An agent here means an AI that does a task step by step on its own, using the apps you give it. Sol at its highest setting scored 33.2% at $0.27 per task. Claude Opus 5 scored 26.9% at eleven times the cost.

Professional work. On Agents' Last Exam, which covers long tasks across 55 sub-industries of computer-based work, Sol scored 56.4%, above Claude Opus 5's highest score at 60% lower cost per task.

Code. On DeepSWE, a test of long software tasks in real codebases, Sol scored 68.8%, within 1.1 points of Claude Fable 5's best, at about 80% lower cost per task. Luna, the cheap one, scored 66.6%, comparable to Claude Opus 5 at medium effort, at 93% less per task. For the cheapest model in the family, that is the surprise of the launch.

Using a computer. On OSWorld 2.0, which has the AI click through real software to finish everyday and professional tasks, Sol matched Claude Opus 5, 60.5% against 60.3%, at about 80% lower cost. OpenAI says Astra remains its best model for this kind of work.

What this means for you: the pattern is the same everywhere. The new models reach roughly the level of last season's top models at a fraction of the cost per finished task.

What got better: fewer mistakes, clearer answers

Fewer factual mistakes. An answer is only useful if the facts are right. OpenAI tested the models on real, anonymised ChatGPT conversations in which users had flagged a factual error by an earlier model. On that test, Sol makes about half as many mistakes as its predecessor, close to Astra's reliability at a much lower cost. OpenAI points out that these conversations were chosen because they went wrong, so errors are rarer in ordinary use.

Clearer, shorter answers. OpenAI brought Astra's way of writing to both new models. It promises "more clarity, less jargon, fewer odd turns of phrase, fewer low-value details, and slightly shorter answers overall without losing substance." Anthropic made almost the same promise for Claude Opus 5.5 on the same day. Both companies heard the same complaint from users, and both answered it.

OpenAI's launch page shows the difference with one request. A user asked for a website redesign in a "bento box" style, a grid of tiles, with a page slider in the top-right corner.

The request: "Could we use a Bento Box design style and add a slider between the pages in the top right?"

GPT-5.6 Sol answers

"Absolutely. I can get the bento feel and the top-right sliding page switcher without pulling in React…"

It ends by quoting the prompt it used for the food illustrations, which the user never asked about.

GPT-6 Sol answers

"The updated site is live. Every page now has a playful bento layout, and the top-right page switcher slides between sections. I checked it on desktop and narrow mobile screens, including browser back navigation."

The same request, answered twice. The older model starts by agreeing and adds details nobody needs. The new one reports what changed and says what it checked. Excerpts quoted from OpenAI's launch page, which shows both replies in full.

OpenAI explains why it prefers the new reply: it "doesn't jump to conclusions as quickly", uses less vague language, and "is more forthcoming with what it did and didn't check." That last point matters most at work. An assistant that tells you what it checked lets you decide what you still need to check yourself.

More honest about its work. Both new models also showed "lower rates of misleading claims about their coding work" than the models they replace. OpenAI's tests deliberately set traps that invite an AI to cut corners and claim success. The full results are in OpenAI's system card.

What this means for you: ask Sol for something you would normally have to double-check, such as a summary with numbers or a small fix to a document, and look at whether it tells you what it verified.

Official sources