GPT-6 Sol and Luna: What's New, Pricing, and Alternatives (2026)
GPT-6 Sol and GPT-6 Luna are two new OpenAI models released on September 22, 2026. They run in ChatGPT Work and Codex, and in the OpenAI API. API pricing is $2 input and $10 output per million tokens for Sol, and $0.10 and $0.50 for Luna.
This guide covers what shipped, where each model is available, and what the launch tests say about finished work. It also covers pricing and the alternatives worth testing. Every figure comes from vendor pages read on September 24, 2026.
What OpenAI released on September 22
OpenAI published the announcement on Tuesday, September 22, 2026. The ChatGPT release notes carry the same date, so two official channels confirm the launch.
The two models sit below GPT-6 Astra in the same family. OpenAI introduced Astra earlier in September as "the most intelligent and aligned model in the world." The new pair brings that generation to faster, lower-cost models.
OpenAI says it trained Sol and Luna "with similar methods as GPT-6 Astra." The announcement lists five areas where the advances carry over: professional work, factuality, coding, computer use, and alignment.
The positioning is clear about the ceiling. OpenAI writes that "GPT-6 Astra continues to be our best model across the board." GPT-6 Sol is the middle option for everyday demanding work, and Luna is the high-volume option.
For developers, the model IDs are gpt-6-sol and gpt-6-luna. For ChatGPT users, the models appear in the model picker inside Work and Codex.
Where you can use Sol and Luna
The availability rules are more specific than usual, so they are worth reading closely.
The announcement says both models are available in ChatGPT Work and Codex "for all Plus, Pro, Business, Enterprise, and Edu users." Free and Go users can use GPT-6 Luna in the desktop app. OpenAI adds that "These models are not yet available in Chat."
The release notes repeat the point in plainer terms. "These models are separate from the models available in Chat." If you open a regular conversation, you will not find them there.
That split matters because Work and Chat do different jobs. OpenAI's help center describes Work as "an agent designed for longer, multi-step work and finished deliverables." It lists the jobs as research, analysis, or creating "a document, spreadsheet, presentation, report, or Site."
So GPT-6 Sol is mainly a model for deliverables, not for quick questions. Our earlier explainer on what ChatGPT Work is covers that agent in more detail.
Two practical notes come from the help center. Availability depends on your plan, workspace settings, and rollout access. The same article says Plus plans include GPT-6 Astra in Work and Codex, so Plus users can compare all three models side by side. And OpenAI says the rollout would run gradually through launch day, so some accounts saw the models later than others.
What the launch tests say about finished work
Most model launches lead with coding scores. Three of the tests here are more useful if your job ends in a report, a spreadsheet, or a deck.
AutomationBench: business workflows across apps
OpenAI describes AutomationBench as "a test of business workflows across apps." In it, agents are tested on "end-to-end workflows using 47 tools across sales, marketing, operations, support, finance, and HR."
OpenAI reports that GPT-6 Sol at xhigh effort outperforms Claude Opus 5 at max effort on this test. It also reports that Luna improves on its predecessor by 5.4 percentage points at high effort.
This is the closest test to real office work in the announcement. A workflow that touches a CRM, a spreadsheet, and an email tool is how many weekly reports actually get built.
Agents' Last Exam: long professional tasks
OpenAI says Agents' Last Exam covers "long-horizon, economically valuable tasks spanning 55 sub-industries." GPT-6 Sol at max effort scores 56.4% on it. The test is useful as a rough signal of how far an agent gets on a multi-hour task before it needs help.
OSWorld: operating a computer
On OSWorld 2.0 offline, GPT-6 Sol at xhigh effort scores 60.5%. OpenAI says that is similar to Claude Opus 5 at medium effort, which scored 60.3%. OpenAI still describes Astra as the world's best model for computer use.
Treat all three as vendor-run results. OpenAI notes that its evaluations ran "in our research environment or via our API," which can differ from production ChatGPT.
Pricing
The API prices below are the amounts OpenAI shows on the announcement page. Prices are per million tokens.
| Model | Input | Output | Where to use it |
|---|---|---|---|
| GPT-6 Sol | $2 | $10 | ChatGPT Work, Codex, OpenAI API |
| GPT-6 Luna | $0.10 | $0.50 | ChatGPT Work, Codex, OpenAI API, and the desktop app on Free and Go |
| GPT-6 Astra | Not listed on this page | Not listed on this page | ChatGPT Work and Codex on Plus and higher plans |
Inside ChatGPT, the cost works differently. The help center says Work "follows the same usage structure as Codex." Your plan includes an allowance, and heavier tasks use more of it.
Signing in to Codex with a ChatGPT account uses your plan's usage and billing. Using your own API key uses API pricing instead. Teams that run both should decide which route each workload takes before usage grows.
Factuality and writing style
Two quieter changes may matter more than the benchmark scores for anyone who publishes numbers.
Fewer factual mistakes. On OpenAI's internal factuality test, GPT-6 Sol "makes about half as many mistakes as its predecessor." The test uses real conversations where users had flagged an error. OpenAI adds a fair caveat: these conversations "are not representative of typical usage, where factual errors are more rare."
Shorter, clearer answers. OpenAI says the new models carry over Astra's communication style. It promises "more clarity, less jargon, fewer odd turns of phrase, fewer low-value details, and slightly shorter answers overall without losing substance."
For report writing, the second change is easy to underrate. A draft that states its findings plainly needs less editing before it goes to a manager.
Caching changes for long agent runs
Alongside the models, OpenAI changed how prompt caching works for the GPT-6 family. The announcement says the new system delivers higher cache hit rates by default.
Developers also get more control. Explicit breakpoints let them choose where a cached prompt prefix ends. Reasoning effort and tool settings can now change between requests without breaking the cache.
OpenAI cites one early result. GitHub reports that the improvements "reduced the share of prompt tokens requiring fresh processing by more than 50%" across billions of requests. Agents resend the same instructions and tool definitions on every turn, and that repeated context is what the cache is built to reuse.
Sol, Luna, or Astra: which one to pick
The family now covers three price and capability tiers. A simple split looks like this:
- Use GPT-6 Sol for recurring reports, analysis write-ups, and spreadsheet-to-deck work inside Work.
- Use Luna for high-volume tasks such as classification, extraction, and first-pass summaries.
- Use Astra when Sol falls short on your own tests, or when the task is long and high-stakes.
A reliable way to choose is to test on your own files. Run one real task through Sol and Luna with the same inputs. Record the time, the cost per task, and every number you had to correct.
Alternatives worth knowing
Two other model releases landed in the same week. Both are worth adding to the same test.
| Model | Released | API price per million tokens | Where to use it |
|---|---|---|---|
| GPT-6 Sol | September 22, 2026 | $2 input / $10 output | ChatGPT Work, Codex, OpenAI API |
| GPT-6 Luna | September 22, 2026 | $0.10 input / $0.50 output | ChatGPT Work, Codex, OpenAI API |
| Claude Opus 5.5 | September 22, 2026 | $4 input / $20 output | Claude API |
| Grok 4.7 | September 21, 2026 | $2 input / $6 output, below 200k prompt tokens | xAI API, Cursor, Grok Build |
Claude Opus 5.5. Anthropic released it on the same day. Its pricing page describes the model as a "Daily driver for agentic coding and enterprise work." Paid Claude plans also list Claude Design, Slides, and Docs as features, which suits deck and document work.
Grok 4.7. SpaceXAI's release notes say it is available on the xAI API as grok-4.7. Pricing is $2 input and $6 output below 200k prompt tokens, and $4 and $12 above. A faster variant is available only through Cursor and Grok Build.
If your work is mostly coding, the benchmark pages for all three are the place to start. If it is mostly reports and decks, a bake-off on last quarter's files will tell you more.
Where a cheaper model stops helping
A lower price per token makes it easier to run more tasks. It does not remove the need to check the output. A report with one wrong figure is still a wrong report, whatever it cost to produce.
That check is where a workspace matters more than the model name. Powerdrill Bloom describes itself as "The AI data analyst that shows its work." Its homepage adds: "Every number comes back with the page, the row and the figure behind it."
The job shape is the same one GPT-6 Sol targets in Work. You start with a spreadsheet or a set of documents and end with a finished file. Our Excel to PowerPoint page covers that route for your own files. For background on the top model in this family, see our GPT-6 Astra write-up.
Pick the model by testing it. Then keep the source trail next to the output, so a reviewer can confirm each number in seconds.
How to start with GPT-6 Sol and Luna
If you use ChatGPT, open Work rather than a regular chat. Choose the model from the picker, then give it one real file and one real deliverable. A quarterly summary or a weekly metrics deck makes a better test than a toy prompt.
If you build on the API, switch the model ID to gpt-6-sol or gpt-6-luna and rerun your evaluation set first. Compare cost per task, not just cost per token. A model that finishes in fewer steps can cost less overall.
If your output is a recurring report, keep a short list of the figures that matter most. Check those first on every run, whichever model produced them.
When you want a spreadsheet turned into a sourced chart, report, or deck, you can try Powerdrill Bloom on your own files.
Frequently asked questions
What are GPT-6 Sol and GPT-6 Luna?
They are OpenAI models released on September 22, 2026, below GPT-6 Astra in the same family. GPT-6 Sol targets demanding everyday work, and Luna targets high-volume tasks. Both run in ChatGPT Work, Codex, and the OpenAI API.
How much does GPT-6 Sol cost?
On the API, OpenAI lists GPT-6 Sol at $2 per million input tokens and $10 per million output tokens. Luna is $0.10 input and $0.50 output. Inside ChatGPT, access comes through your plan's Work and Codex allowance.
Is GPT-6 Sol available in regular ChatGPT chats?
No. The release notes say Sol and Luna "are separate from the models available in Chat." You find them in Work and Codex. Free and Go users can use Luna in the ChatGPT desktop app.
What is the difference between GPT-6 Sol and GPT-6 Astra?
Astra is OpenAI's top model, which it calls its "best model across the board." Sol was trained with similar methods and targets faster, lower-cost work. OpenAI suggests Astra for the most demanding and important projects.
Why use GPT-6 Luna instead of Sol?
Luna costs less per token and suits high-volume, simpler tasks such as extraction and first-pass summaries. OpenAI reports that Luna at higher effort matches GPT-5.6 Sol on its factuality test. Sol is the better fit for complex, multi-step deliverables.
Sources: OpenAI, Introducing GPT-6 Sol and Luna · ChatGPT release notes · ChatGPT Work and Codex help article · Claude pricing · xAI API release notes. Prices and availability read from vendor pages on September 24, 2026.