OpenAI is launching GPT-6 Sol and GPT-6 Luna with a promise that is less spectacular than a new giant model, but potentially more significant for everyday use: making the GPT-6 family much cheaper to use.
GPT-6 Soil costs 2 dollars per million tokens in entry and 10 dollars in release into API. GPT-6 Luna goes down to 0,10 dollar in entry and 0,50 dollar out. OpenAI states that these rates represent a decrease of 50 % from GPT-5.6 Sol and Luna's promotional rates.
For creators, consultants, freelancers and small businesses, the signal is clear: the next step of the AI is not just to get a smarter model. It consists of choosing the level of intelligence that is actually needed for each task in order to automate more without exploding costs.
Information verified on 23 September 2026. OpenAI indicates that Sol and Luna are available today in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users. Free and Go users can access GPT-6 Luna in desktop application. Deployment in ChatGPT is progressive during the day and both models are available in the API under the names gpt-6-ground and gpt-6-luna.
GPT-6 Sol and Luna: what OpenAI is actually announcing
OpenAI had launched GPT-6 Astra at the beginning of the month as his most capable model for complex tasks, computer use, research and professional workflows from end to end. Sol and Luna do not replace Astra. They complement the GPT family-6 with a different compromise between intelligence, speed and cost.
OpenAI presents GPT-6 Soil as the model for hard daily work that requires more reasoning, but for which use Astra at each request would be disproportionate. GPT-6 Luna targets high volume, faster and much cheaper tasks.
This segmentation is important. For several years, the reflex was to look for the most powerful model available. With agents that can execute tens or hundreds of steps, the cost per action becomes as important as gross performance.
In the API, GPT-6 Sol moves to 2 dollars per million input tokens and 10 dollars for output. GPT-6 Luna costs 0.10 dollar for input and 0.50 dollar for output. OpenAI retains Astra as its best overall model and recommends reserving it for tasks where maximum quality justifies its cost.
Why this price drop matters more than it seems
A one-time conversation consumes relatively little resources. Agent workflow is different. The agent can read a history, search for multiple sources, call tools, go back on his results, produce a draft, correct it and trigger another step. A single mission can therefore generate many calls to the model.
OpenAI also highlights an improvement of the quick hiding. The principle is simple: when an application reuses a large part of the same context, it does not need to completely reprocess this information at each call. Cached inputs can benefit from a reduction of 90 %. OpenAI also states that GitHub has reduced by more than 50 % the share of tokens of prompts requiring new processing on billions of queries thanks to these improvements.
For a small business, this changes how to calculate the profitability of automation. The question is no longer only: "What AI gives the best answer?" It becomes: "What quality is needed here, how many times will this task be carried out and what model offers the best relationship between result and cost? »
VIFLY’s perspective: iA professional enters a phase where knowing how to choose the right model becomes as important as knowing how to write a good quick. Using the most powerful model for each micro-task is the digital equivalent of using a senior consultant to copy data.
Two concrete examples for users VIFLY
A creator who transforms a video into multiple content
Imagine a creator who publishes a long video every week. Its workflow includes transcription, identification of highlights, five post proposals, a newsletter, social descriptions and a summary of comments.
Transcription, comment ranking or format transformation can be entrusted to a fast and economical model such as Luna. Editorial analysis, the selection of the main angle or the writing of a strategic article may justify Sol. An exceptionally complex mission, which crosses many documents or has to produce a high-level deliverable, can remain reserved for Astra.
This architecture keeps the human where its value is highest: the point of view, experience, validation and relationship with the audience. The LinkHub VIFLY can then centralize useful destinations from these content instead of dispersing the audience between several links.
A consultant who prepares his appointments automatically
A consultant receives several appointments a week. For each, a workflow can retrieve the information provided by the prospect, summarize its need, classify the type of mission and prepare an appointment sheet.
Much of this preparation does not systematically require the most expensive model. Luna can handle repetitive operations, Sol can prepare the synthesis and relevant questions, then the consultant valid before the appointment. If making an appointment goes through VIFLY Booking, the route is already more structured: availability, service and reservation become usable data rather than a suite of informal messages.
The gain is not only from a cheaper model. It comes from the possibility of building a workflow where each step uses the appropriate level of intelligence.
Sol, Luna or Astra: stop choosing a theme based on prestige
GPT-6 Astra remains the model that OpenAI positions for the most demanding projects. Sol aims at a more accessible balance between capacity and cost. Luna is more focused on speed, volume and economy.
OpenAI publishes several evaluations showing that Sol is making significant progress on professional workflows, news, code and computer usage. On AutomationBench, which tests end-to-end workflows through 47 tools in areas such as marketing, sales, operations and finance, GPT-6 Sol gets according to OpenAI 33,2 % at xhigh effort for a cost of 0,27 dollar per task in the measured configuration.
These benchmarks are useful for comparing trends, but they must not become absolute arguments. A benchmark score does not capture your activity, documents, constraints or data quality.
The right test is to take a real repetitive task and measure three things: time saved, error rate and total cost. If Luna correctly realizes 90 % a workflow for a fraction of the price, use Astra by default does not necessarily bring value. If a complex analysis influences an important business decision, saving a few cents at the expense of quality can be a poor optimization.
Nuances and points to bear in mind
First grade: the prices announced here concern API. The experience and limitations in ChatGPT depend on the subscriptions and quotas applied by OpenAI. Therefore, it is not necessary to turn a decline in API price into a promise to lower ChatGPT subscription prices.
Second shade: OpenAI states that deployment in ChatGPT is progressive during the day of 23 September. The absence of Sol or Luna in an account at the time of reading does not necessarily mean that this account is excluded.
Third shade: the models are not yet available in the same way on all surfaces. OpenAI explicitly indicates their availability in ChatGPT Work and Codex for eligible plans, Luna in desktop application for Free and Go, and a progressive deployment in ChatGPT. The situation can change rapidly.
Finally, a cheaper model can encourage more automation. This should not lead to more permission. An AI responsible for preparing an email does not necessarily need to be able to send it. An AI that analyzes reservations does not need to change payments. The lower cost of execution does not reduce the need for human validation.
What should we do next? Develop a proper model strategy
- List three repetitive tasks. Choose tasks you perform each week: synthesis, content preparation, demand qualification or data analysis.
- Sort them by level of risk. Reformulation is low risk. A commercial proposal or action on a customer account is more sensitive.
- Test the least expensive model that can succeed. Start low, then go up in range only when the quality requires.
- Measure the actual result. Time, necessary corrections, cost and final quality are worth more than a general benchmark.
- Keep human validation at critical points. Publication, payment, deletion, contractual commitment and sensitive customer communication must remain monitored.
- Structure your course. An AI works better when the information and actions are clear. VIFLY allows to centralize its presence, while LinkHub organizes the destinations and Booking structure the reservation.
This method avoids two excesses: paying too much for simple tasks and undersized the model on tasks that really matter.
The real change: intelligence is becoming a resource to be allocated
GPT-6 Sol and Luna are less interesting because they add two names to the model selector than because they illustrate a new AI economy.
When models cost less, companies can integrate them into more steps. As the caching progresses, long conversations and agents who reuse context become less expensive. And when several levels of models coexist, orchestration becomes a skill.
For the independents, this can be excellent news. Large companies have long benefited from teams capable of automating operations. A small structure can now gradually build its own workflows, provided that it does not confuse automation with the abandonment of control.
The VIFLY strategy remains the same: use the AI to reduce friction between an intention and a result, while keeping human value where it counts. On the blog VIFLYtherefore, the goal is not to follow each new model as a race of numbers, but to identify what it really changes in daily work.
GPT-6 Astra showed how far the GPT-6 family could go. Sol and Luna ask another question, perhaps more important for common uses: how much intelligence does it really take to do a job properly?
FAQ – Frequently Asked Questions
What is the difference between GPT-6 Astra, Sol and Luna?
Astra remains the most capable model of the family according to OpenAI. Sol aims at a balance between advanced performance and cost, while Luna favours speed, volume and a very low price.
How much does GPT-6 Soil in API?
At 23 September 2026, OpenAI displays 2 dollars per million tokens and 10 dollars per million tokens for GPT-6 Sol.
How much does GPT-6 Luna cost?
OpenAI displays 0,10 dollar per million tokens of entry and 0,50 dollar per million tokens of exit in API.
GPT-6 Are Sol and Luna available in France?
OpenAI does not announce in its press release a specific restriction on France. The models are available in the indicated surfaces and plans, with progressive deployment in ChatGPT during the day. The exact availability can therefore depend on the account and the product used.
Should we replace Astra with Sol?
No. The choice depends on the task. Astra remains intended for the most demanding needs. Sol can be more rational when its level of performance is sufficient, especially in frequently executed workflows.
Does an independent need several AI models?
Not necessarily at first. But as soon as an activity automates several tasks, using different levels of models can reduce costs while maintaining more power on important steps.