GPT-6 Sol and GPT-6 Luna, the two new models from OpenAI, have been available since September 22, 2026. The company prices them at half the cost of the GPT-5.6 generation. For an SME, this announcement changes one thing above all: the real cost of an AI project.
This article first sums up the facts, then turns them into concrete decisions. It speaks to executives and managers who run a budget, not only to developers.
Key takeaways
- OpenAI launched GPT-6 Sol and GPT-6 Luna on September 22, 2026. GPT-6 Astra nevertheless remains its most capable model.
- Sol costs $2 per million input tokens and $10 per million output tokens. Luna costs $0.10 for input and $0.50 for output.
- Both models accept about 1.05 million tokens of context and produce up to 128,000 tokens per response.
- OpenAI states that Sol makes about half as many factual errors as its predecessor. That figure, however, comes from an internal evaluation.
GPT-6 Sol and Luna: what OpenAI announced
OpenAI now organizes its GPT-6 lineup in three tiers. Astra targets the hardest tasks. Sol targets code and demanding agents. Finally, Luna handles focused, high-volume work.
The two new models replace GPT-5.6 Sol and GPT-5.6 Luna. Developers access them through the API. Plus, Pro, Business, Enterprise, and Edu subscribers, for their part, use them in ChatGPT Work and in Codex. According to OpenAI, users on the Free and Go plans access Luna in the desktop app.
Both models read text and images, and they respond in text. The company also highlights a clearer style, with less jargon.
The pace of releases remains fast. On September 29, 2026, at its DevDay, OpenAI already presented GPT-6.1 Sol. A recent model therefore stays recent for only a short time, and a company must factor this into its technical choices.
Prices cut in half: how to read the figures
A token is a fragment of a word. Providers bill separately for the text sent to the model (the input) and the text it produces (the output). OpenAI announces a 50% drop compared with the equivalent GPT-5.6 versions.
- GPT-6 Sol: $2 for input and $10 for output, per million tokens.
- GPT-6 Luna: $0.10 for input and $0.50 for output, per million tokens.
- Cache: OpenAI grants a 90% discount on input tokens that the model rereads from the cache.
The cache matters to companies that often resend the same context: a catalog, a procedure, a customer history. The model then rereads this content at a reduced rate, which lowers the bill for assistants and agents.
Beware of the shortcut, though. The token price does not tell you the cost of a task. Both models offer a reasoning effort setting, from the minimum level to the maximum level. The higher the effort, the more tokens the model consumes.
The specialized outlet DataNorth also points out another limit. OpenAI has published no test that shows how the two models hold up when the context approaches the maximum. A very long document therefore does not guarantee a reliable answer.
Why this announcement matters for an SME
Lower prices move the line of what is profitable. Uses that were too expensive yesterday now become affordable: reading every incoming email, checking every supplier invoice, enriching every product listing.
The two-speed lineup also simplifies trade-offs. Luna suits sorting, extraction, and summarization in large quantities. Sol then takes over when the task requires several steps, tools, and judgment.
Reliability is improving, according to the company. Keep one reservation, however: OpenAI measures this progress on its own test sets, and Requesty notes that these conversations do not represent ordinary use. A human must therefore keep validating the outputs that commit the company.
How to take advantage of GPT-6 right now
- Target three repetitive tasks. Choose the ones that take up the most hours: data entry, standard replies, meeting reports.
- Test the small model first. Then keep the cheapest model that passes your own test cases.
- Measure the cost per task. Count the tokens of a complete task, retries and corrections included.
- Stay independent of the model. A well-designed architecture switches models without rewriting the application.
- Set rules for data. In particular, check the contractual terms before sending customer or financial data.
Train your teams too. An employee who knows how to write an instruction and verify an answer gets more value from the same model. Our training courses for teams can support this skills development.
The AISYSNEXT point of view
At AISYSNEXT, we treat the model as a component. The value comes instead from integrating it into your data and your processes. A cheaper model fixes neither scattered data nor an unclear process.
We recommend a three-step approach: an audit of tasks, a limited pilot, then a measured rollout. Our team for integrating artificial intelligence into your business supports this scoping. When a standard tool is not enough, our custom development team builds the business application that connects the model to your systems.
You can also find our other analyses of AI models and agents on the AISYSNEXT blog.
Frequently asked questions
GPT-6 Sol or GPT-6 Luna: which one should you choose?
Start with Luna for simple, high-volume tasks. Then move to Sol if your tests reveal errors on complex cases. Astra remains the most capable option: reserve it for needs that justify it.
Can you use GPT-6 for free?
According to OpenAI, Free and Go users access Luna in the desktop app. Sol, on the other hand, requires a paid subscription or API access, which the company bills by usage.
Should you migrate your tools to GPT-6 right away?
There is no rush. First replay your real use cases on the new model. Compare quality, turnaround time, and cost per task, then decide.
Want to put a number on a first AI use case in your company? Request a free quote: our team studies your need, then replies with a clear proposal.
