GPT-6 Sol and Luna: why AI got half as expensive

On the same Tuesday, September 22, 2026, two of the largest AI companies in the United States announced the same thing in different words: you can pay less. OpenAI launched GPT-6 Sol and GPT-6 Luna at half the price of previous models. On the same day, Anthropic lowered the price of Claude Opus, the company's heavy-duty AI.

Days earlier, Anthropic's chief had asked the sector to slow down. And, on the other side of the world, Chinese models that anyone can download for free charged a fraction of these prices for use.

Here you will understand what GPT-6 Sol and Luna are, how much they really cost, why AI is getting cheaper, who is pressuring this cut, and what changes for you.

Quick summary: the answer in 30 seconds. GPT-6 Sol is OpenAI's daily work model, and GPT-6 Luna is the fast and cheap one. Both cost half of previous versions for those who pay per use. The price dropped due to efficiency, a memory system that avoids rereading the same text, and competition from cheap Chinese models. For those who just chat with ChatGPT, little changes; for companies and programmers, the bill can drop by half.

Index

• What are GPT-6 Sol and GPT-6 Luna?
• How much does GPT-6 cost: did the price really drop by half?
• Why is AI getting cheaper?
• GPT-6 Sol or Claude Opus 5.5: which is cheaper?
• And Google? Gemini's low price has an expiration date
• The Chinese AI models pressuring prices
• Does asking to slow down AI and lower prices make sense?
• What changes for you
• Frequently asked questions

What are GPT-6 Sol and GPT-6 Luna?

GPT-6 Sol and GPT-6 Luna are the two cheapest models of OpenAI's new GPT-6 generation, which currently has three models. Sol is the "everyday car" for heavy work, like programming and analyzing data. Luna is the economical one, made for short, repeated tasks, like summarizing documents, extracting information from text, or answering quick questions.

A language model is the "brain" that responds when you chat with ChatGPT. It's a program trained with an enormous amount of text to predict which word comes after another, like a phone's autocorrect millions of times more sophisticated.

OpenAI doesn't have just one model. It maintains a ladder: the higher the step, the more capable the model and the more expensive each use. In early September, the company launched the top of the ladder, GPT-6 Astra, which we explained in our article about GPT-6 Astra and why it scares so many people. Now came the lower steps.

The previous generation, GPT-5.6, still had a middle step, Terra. In GPT-6, it hasn't been announced yet.

In the table below, the price is charged per token, a piece of a word. Think of arcade tokens: each piece of text that goes in or out spends a token. We'll explain better further on.

Prices from OpenAI's official announcement and the GPT-6 Astra page, checked on 09/26/2026. Input is the text you send; output is what the model responds.

According to OpenAI, Sol and Luna were trained with methods similar to Astra's. The promise is to bring part of the top model's intelligence to faster and cheaper models.

The company says that, in a test with real conversations where users marked errors, GPT-6 Sol makes about half the errors of its predecessor. Luna, when you ask it to think more before responding, gets as much right as the old GPT-5.6 Sol, costing about 1% of what GPT-5.6 Sol charged for the same task. These are numbers measured by OpenAI itself.

Where to use GPT-6 Sol and Luna today

Both models are in ChatGPT Work, the ChatGPT area focused on work tasks, and in Codex, OpenAI's programming tool. They're available to Plus, Pro, Business, Enterprise, or Edu subscribers. Those using the free plan or Go access GPT-6 Luna in the ChatGPT desktop app.

A confusing detail: according to OpenAI, the models aren't yet in Chat, the common ChatGPT conversation mode. Subscribers need to open ChatGPT Work or Codex to use Sol. In the API, the "socket" that connects other programs to the model, they appear as gpt-6-sol and gpt-6-luna.

In one sentence: Sol is the heavy-duty model and Luna is the economical one, and both sit below Astra on OpenAI's ladder.

How much does GPT-6 cost: did the price really drop by half?

It dropped, but it depends on the comparison. GPT-6 Sol costs $2 per million input tokens and $10 output, half of what OpenAI charged for GPT-5.6 Sol in September. But that old price was already promotional. Compared to the full price of GPT-5.6 Sol, the cut exceeds 60%.

Explaining: the official GPT-5.6 Sol page shows $4 for input and $20 for output as a promotional price, "at least until November 21, 2026." The same page says this was a 20% discount on input and 33% on output. Calculating backwards, the full price was $5 and $30.

Prices per million tokens, from OpenAI, checked on 09/26/2026.

Luna has another story. In July, OpenAI had already cut 80% from the price of GPT-5.6 Luna, according to that time's announcement. Now it dropped again. In two months, the company's cheapest model dropped in price twice.

Another difference: the GPT-5.6 Sol page warned that the price was promotional, with a deadline. The GPT-6 announcement doesn't mention promotion or deadline. This matters to companies, which need to plan a product without fear of the price rising when a promotion ends.

What a token has to do with your bill

A token (piece of a word) is the unit models use to read and write. A short word is usually one token; a long one becomes two or three pieces. Think of arcade tokens: each piece of text that goes in or out spends a token.

Those who use ChatGPT through the app don't see these tokens: they pay the subscription, or nothing, and have a usage limit. Those who create a product on top of the model use the API and pay per token, like an electricity bill. It's for this audience that the price per million tokens changes everything.

An example of ours, just to get an idea of the bill's size. Imagine a store that uses AI to answer a thousand customer questions per day. Each question sends 2,000 tokens (the history and the store's rules) and receives 500 in response.

Calculation with official prices from 09/26/2026, without cache discount. In real life, the cheapest model only works if it gets the answer right.

In one sentence: GPT-6 dropped to half price, but half of a price that was already on promotion, and Luna got even cheaper than it already was.

Why is AI getting cheaper?

For three reasons that add up. Servers became more efficient at running the models. Cache made it cheaper to reread what the model already read. And new models spend fewer tokens to complete the same task. Behind it all, there's pressure from competitors that charge much less.

OpenAI itself says that improvements in cache and inference allowed serving the models for less, and that it's passing on these savings. It's worth understanding both terms.

Inference: the cost of each response

Inference is the work the computer does every time the model responds. Training a model is like writing a recipe book; inference is cooking each dish a customer orders. The book is written once. The dishes come out millions of times a day.

When the company learns to cook faster, with less gas and fewer pans, each dish comes out cheaper. This is what companies call efficiency gains in inference.

Cache: not paying twice for the same reading

Cache (short-term memory) stores text the model already read so it doesn't have to process it again. Think of a teacher grading a hundred essays on the same topic: they read the prompt once and then just consult it.

This weighs heavily on AI agents, programs that don't just respond but act on their own for hours. An agent rereads the same material hundreds of times: the instructions, the files, the history. If you want to understand this type of program, see our guide on agentic AI.

OpenAI says GPT-6 gives a 90% discount on rereads stored in cache. According to the company, GitHub, where programmers store and share code, reported that with these improvements, Copilot, GitHub's programming assistant, now processes from scratch less than half the text it processed before.

Anthropic tells the same story with its own numbers. In the Claude Opus 5.5 announcement, the company says cache rereads are most of the cost in programming and agents, and lowered this price by 60%.

Models that spend fewer tokens

The third reason is more subtle. A new model can charge the same per token and still come out cheaper, because it solves the task with less text. It's like a car that, besides using cheaper gasoline, goes more miles per gallon.

Anthropic gives the clearest example: the price per token of Opus 5.5 dropped 20%, but the bill for typical work drops about 40%. The difference comes from the model spending fewer tokens to reach the same result.

In one sentence: AI got cheaper because it became more efficient to run, remembers what it already read, and talks less to do the same, and because competition tightened.

GPT-6 Sol or Claude Opus 5.5: which is cheaper?

Per token, GPT-6 Sol costs exactly half of Claude Opus 5.5: $2 versus $4 on input and $10 versus $20 on output. But price per token isn't price per completed task, and so far no one has compared the two models under the same conditions.

Claude Opus 5.5 is Anthropic's heavy-duty model, launched the same day. It costs 20% less per token than Opus 5 and 60% less than Fable 5.1 and, according to the company, works at the level of the top model, Fable 5.1, on most tasks. We covered everything about it in the article Claude Opus 5.5: Fable's power paying 60% less.

Official prices from OpenAI and Anthropic, checked on 09/26/2026. On cache rereads, both tie.

The American tech site VentureBeat summarized the trap: a cheaper model can cost more if it needs more attempts, generates more text, or makes mistakes more frequently. Each extra attempt is a new bill.

OpenAI published that Sol, at maximum effort, solves 33.2% of tasks in a business automation test, costing $0.27 per task. In the same measurement, Claude Opus 5 solved 26.9%, costing 11 times more per task. But the comparison was with Opus 5, not Opus 5.5, which came out hours earlier. And it used numbers from other companies, measured under different conditions.

In one sentence: Sol is cheaper per token, Opus 5.5 promises to spend fewer tokens, and only a side-by-side test on your tasks decides who wins the bill.

And Google? Gemini's low price has an expiration date

Google charges even less for its intermediate model, Gemini 3.8 Flash: $0.75 per million input tokens and $3.75 output. But this is a launch price, valid until December 31, 2026. Starting January 1, 2027, it doubles.

The numbers are in the official Gemini 3.8 Flash announcement, checked on 09/26/2026. Even after the increase, Flash, at $1.50 and $7.50, remains cheaper per token than GPT-6 Sol.

The difference is in strategy. The GPT-6 Sol announcement doesn't mention a deadline for the price. Google makes clear its discount is temporary. For a company planning a product for next year, this changes the calculation.

The same week, Google bet on another front: Gemini 3.8 Live with Live Avatar, an animated face that speaks 97 languages, for now only for businesses on Gemini Enterprise.

In one sentence: Gemini 3.8 Flash costs much less than GPT-6 Sol, but the low price ends in December and then doubles.

The Chinese AI models pressuring prices

Open-weight Chinese models are the background pressure in this race. Companies like Xiaomi and DeepSeek launch models that approach American ones in various tests, charge a fraction of the price, and still let anyone download the model for free.

The American site Ars Technica summarized the movement: OpenAI and Anthropic are racing to compete with open models, because companies started using expensive models only when needed and sending the rest to cheap alternatives.

An open-weight model is one whose ready "brain" is available for download. Think of the difference between buying a dish at a restaurant and receiving the complete recipe: with the recipe, you cook at home and don't pay the restaurant anymore. The company can run the model on its own computers, without paying per token.

Xiaomi, better known in Brazil for phones, launched the MiMo-V2.6 family the same week. MiMo-V2.6-Pro scored 46 points on Artificial Analysis's intelligence index, an independent company that compares models. It was first place among open models of similar size, in a group where the middle score is 18.

Xiaomi's models are on Hugging Face, the largest open model repository, with an MIT license. It's one of the freest licenses that exist: it allows using, modifying, and selling products made with the model.

Prices per million tokens, from the official pages of Xiaomi, DeepSeek, OpenAI, Google, and Anthropic, checked on 09/26/2026.

The table shows why Luna exists. It's OpenAI's model that competes head-to-head with Chinese ones on price: it charges less on input than MiMo Flash, but almost double on output. Sol costs about eleven times MiMo-V2.6-Pro on output.

DeepSeek has an extra trick: it charges half price during off-peak hours defined by the company, like cheaper electricity rates at dawn. During this time, cache rereads cost $0.003 per million tokens.

Not everything is an advantage. Running a Chinese model through the company's own API means sending your data to servers in China, something many companies won't accept. And running the model on your own computers requires expensive machines and people to maintain them.

In one sentence: open Chinese models cost a fraction and can be downloaded for free, and GPT-6 Luna is the most direct American response to them.

Does asking to slow down AI and lower prices make sense?

It makes more sense than it seems. The request to slow down was about the capability of the most advanced models, not about price. And the September 22 launches didn't bring intelligence leaps: they brought models similar to previous ones, for much less money.

In September, Dario Amodei, Anthropic's chief, published the essay We Must Pace the Frontier. In it, he wrote that we need to "slow down the pace at which we improve AI models' capabilities," so that risk prevention has time to keep up. We covered this story in the article about AI safety and the request to slow down.

Amodei himself left an opening in the text: slowing down doesn't mean stopping training or technical progress. It means giving time to ensure models do what people want, without causing harm, and for outside evaluators to confirm they're safe.

Ars Technica gave the most accurate headline of the week: both launches promise "a little more for a lot less money." For the site, the AI race entered the price research phase.

The analysis goes further. According to Ars, many companies no longer want a smarter model: they want what already exists to work predictably and at a reasonable price. This demand shift could become a natural slowdown, without anyone needing to decree it.

In one sentence: Anthropic's chief asked to slow down model intelligence, not price, and the week showed the fight now is over who delivers more work per dollar.

What changes for you

For those using ChatGPT for free, little: GPT-6 Luna arrives through the desktop app. For subscribers, Sol arrives at ChatGPT Work and Codex. The big change is for those who pay per use: programmers and companies can cut their bill in half.

Our opinion: the price cut is good news for those building products with AI, because cost stops being an obstacle for many small projects. But performance numbers are still from the companies themselves. Before switching an entire process, test with your own tasks.

If you want to follow the next chapters of this dispute, articles in our artificial intelligence section cover each launch.

Frequently asked questions

Is GPT-6 Sol free?

No. GPT-6 Sol is in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, or Edu subscribers. Those using the free plan have access to GPT-6 Luna, but only in the ChatGPT desktop app.

What's the difference between GPT-6 Sol and GPT-6 Luna?

Sol is the heavy-duty model, made for programming, code review, and data analysis. Luna is the economical one, for short, repeated tasks, like summarizing and extracting information. In the API, Luna costs one-twentieth of Sol's price, both on input and output.

Did GPT-6's price really drop by half?

It dropped by half compared to the price OpenAI charged in September for GPT-5.6. For Sol, that old price was promotional; from the full price, the drop exceeds 60%. For Luna, the output drop reaches 58%.

What is an open-weight AI model?

It's a model whose ready "brain" is available for download. The company or programmer can run the model on their own computers, without paying per use. The Chinese MiMo, from Xiaomi, and DeepSeek are examples.

Is GPT-6 Sol better than Claude Opus 5.5?

Can't say yet. Sol costs half per token, but no one has compared the two under the same conditions. Tests published by OpenAI compare Sol with Opus 5, Claude's previous version.

---

The week's message is that good AI got cheap, and the next step in the dispute is proving who delivers more finished work per dollar. If you use AI at work, the best way to know is to test Sol and Luna on your own tasks.