In short: OpenAI released GPT-6 Sol and GPT-6 Luna on September 22, 2026. They are cheaper, faster members of the GPT-6 family, trained with the same methods as GPT-6 Astra. Sol is for coding and agent work; Luna is the budget model for quick, high-volume jobs. Both cost half what the GPT-5.6 versions did. The catch is that neither is in regular ChatGPT chat yet. Paid users find them in ChatGPT Work and Codex, and free users can try Luna only in the desktop app.
Update, 30 September 2026: OpenAI has released GPT-6.1 Sol, an upgrade to Sol at the same price: $2 per million input tokens and $10 per million output tokens, with cached input cut to $0.10. OpenAI says it nearly matches GPT-6 Astra on coding, computer use and professional work. Like Sol, it is in ChatGPT Work and Codex for paying users and not yet in regular chat. The figures below are for the original GPT-6 Sol. Google’s new top model launched a day later at the same introductory price: see our Gemini 4 Argon guide.
OpenAI announced these two on the same afternoon Anthropic shipped Claude Opus 5.5, so it’s easy to assume they are OpenAI’s answer to Opus 5.5. They are the cheaper options; Astra is still OpenAI’s top model. Below is where to find them and what they cost for a real task.
Can you use GPT-6 Sol or Luna today?
Probably, just not in the regular chat window. OpenAI’s availability note says both models are “not yet available in Chat.” Here is who gets what:
- Plus, Pro, Business, Enterprise and Edu: both models, inside ChatGPT Work and Codex.
- Free and Go: GPT-6 Luna only, and only in the ChatGPT desktop app.
- Developers: both in the OpenAI API as
gpt-6-solandgpt-6-luna.
OpenAI is rolling them out gradually through the day, so if the picker doesn’t show them yet, check again later. Plan prices have not changed. Our ChatGPT pricing guide covers what each plan costs.
Sol and Luna do different jobs
OpenAI’s model docs describe Sol as built for “complex coding and agentic workflows.” That means tasks where the model runs for a while, uses tools and makes changes across many files. Luna is meant for “focused, high-volume tasks” like summaries and quick answers.
Both can read about 1.05 million tokens at once (a token is roughly three-quarters of a word, so that is several long books). You can also set how hard the model thinks, from “none” up to “max.” More effort gives better answers on hard problems and costs more.
OpenAI also says the pair picked up Astra’s writing style, with less jargon and slightly shorter answers. OpenAI’s launch examples back this up, but that’s their claim, not our test.
What does GPT-6 Sol cost next to Luna and Claude?
API prices per million tokens are below, with Astra and Claude Opus 5.5 for scale.
| Model | Input | Output |
|---|---|---|
| GPT-6 Luna | $0.10 | $0.50 |
| GPT-6 Sol | $2 | $10 |
| Claude Opus 5.5 | $4 | $20 |
| GPT-6 Astra | $10 | $50 |
Per-million prices are hard to picture, so I priced one ordinary job: summarizing a 30-page report. That is about 20,000 tokens in and a 1,000-token summary out. My arithmetic from the list prices:
| Model | One report | 1,000 reports |
|---|---|---|
| GPT-6 Luna | $0.0025 | $2.50 |
| GPT-6 Sol | $0.05 | $50 |
| Claude Opus 5.5 | $0.10 | $100 |
| GPT-6 Astra | $0.25 | $250 |
These are minimum costs. When the model thinks harder, that thinking is billed like written output, so higher effort settings raise every row. Look at the gap between rows. Luna does this job for a hundredth of what Astra charges. Cached input (text you have already sent once) also gets a 90% discount on GPT-6, which helps if you send the same long document repeatedly.
If you only use the ChatGPT app, none of this changes your bill. It matters once you build with the API or connect ChatGPT to automation tools like Make.
OpenAI’s benchmarks predate the new Claude
OpenAI’s charts compare Sol and Luna against Claude Opus 5 and Fable 5.1. Opus 5.5 came out the same day, so it is missing. Keep that in mind when you read the headline claims:
- Business workflows (AutomationBench): Sol at its second-highest effort setting scored 33.2%, ahead of Opus 5 at 26.9%, at about 9% of Opus 5’s cost per task. Anthropic’s own table puts Opus 5.5 at 40.0% on the same benchmark.
- Coding (DeepSWE v1.1): Sol scored 68.8% and Luna 66.6% at max effort, close to Claude Fable 5 at 69.9%.
- Computer use (OSWorld 2.0): Sol scored 60.5%, level with Opus 5 at medium effort (60.3%), at about 80% lower cost.
- Factual mistakes: on OpenAI’s internal test, Sol makes about half as many errors as the model it replaces.
Both companies report the same 26.9% for Opus 5 on AutomationBench, so the Opus 5.5 and Sol figures are probably comparable. On that test, Opus 5.5 scores higher and Sol costs less. Either way, Sol and Luna are among the cheapest models you can get for this level of quality right now.
Which one should a beginner pick?
My verdict: if you pay for ChatGPT, open ChatGPT Work and try Sol on something real, like cleaning up a spreadsheet or drafting a multi-step plan. Keep Astra for the jobs where a wrong answer costs you. On the free plan, Luna in the desktop app is a free way to see the GPT-6 writing style, but don’t expect it to replace your regular chat yet.
Building something on the API? Start with Luna and move up to Sol only when Luna’s answers aren’t good enough. Test on your own prompts before you switch anything customers rely on, since cheaper models can fail in ways a benchmark won’t show. The full year of OpenAI launches is in what’s new in ChatGPT 2026.
Questions readers are asking
Does GPT-6 Sol replace GPT-6 Astra?
No. OpenAI says Astra “continues to be our best model across the board.” Sol and Luna replace the GPT-5.6 Sol and Luna models at a lower price.
Why can’t I see Sol in my ChatGPT chat?
They launched in ChatGPT Work and Codex first, and OpenAI says they are not yet in Chat. The rollout is also gradual, so some accounts get them later in the day.
Is GPT-6 Luna free?
Free and Go users can use Luna in the ChatGPT desktop app, within their plan’s usage limits. In the API, every model is pay-per-token.
Are Sol and Luna open-weight models?
No. Open-weight means you can download a model and run it yourself, and you can’t do that with these. Like Astra, they run only on OpenAI’s servers and through partners such as GitHub Copilot.
Get the next launch explained the same day
Our daily newsletter explains each launch and whether it matters to you, in plain English. It’s free.
Sources
- OpenAI: Introducing GPT-6 Sol and Luna (announcement, pricing, benchmarks, availability)
- OpenAI API docs: GPT-6 Sol and GPT-6 Luna (context window, effort levels, cached pricing)
- Anthropic: Introducing Claude Opus 5.5 (Opus 5.5 pricing and AutomationBench score)
- TechCrunch: OpenAI launches GPT-6 Sol and Luna
You may also like
- Gemini 4 Argon: Who Gets It, When
- GPT-6 Astra: What Really Changed
- What’s New in ChatGPT 2026
- ChatGPT Pricing: Every Plan Explained
- Claude Opus 5.5: What Changed
- What Is ChatGPT Codex?
- Every AI Model Worth Knowing in 2026
Two ways to go further
The AI Prompt Library
1,000+ ready-to-use prompts for Claude, ChatGPT, and Gemini. Stop staring at a blank box.
Get it for $39 →2-Hour Live AI Crash Course
A private, beginner-friendly session across Claude, ChatGPT, Gemini, and the wider landscape.
Book for $125 →