GPT-6 Sol and Luna: Prices, Access and Differences
OpenAI officially launched GPT-6 Sol and GPT-6 Luna as lower-cost GPT-6 options for professional and high-volume work across Codex, ChatGPT Work and the API.
GPT-6 Sol and Luna are now official OpenAI models, ending the recent community speculation around the Sol name. OpenAI launched both on September 22, 2026 as lower-cost members of the GPT-6 family. Sol is positioned for complex professional, coding and agent work; Luna has the lower listed API price. They are rolling out in ChatGPT Work and Codex, and their API model IDs are gpt-6-sol and gpt-6-luna. Read OpenAI’s launch announcement.
GPT-6 Sol and Luna: what became official
The release confirms the model labels that developers had been discussing before OpenAI published them. Our earlier GPT-6 Sol rumor record documented that discussion while clearly separating screenshots and routing reports from a verified product. That page is now useful as a dated record of what people expected; this page covers the official launch.
OpenAI describes the new pair as extending GPT-6 capabilities to faster and more affordable workloads rather than replacing GPT-6 Astra at the top of the family. Astra remains the company’s recommended choice when a task calls for its strongest results. Sol occupies the middle decision point: difficult work where quality matters, but repeated attempts and long-running agent activity make cost important. Luna has the lower listed API price, while OpenAI presents Sol and Luna together as ways to make GPT-6 practical for more everyday applications at scale.
This creates a practical three-level GPT-6 lineup. The useful comparison is not that one model is universally best. It is which model delivers enough quality for a defined task while preserving the latency, usage allowance or API budget that the workflow needs.
GPT-6 Sol pricing versus GPT-6 Luna pricing
OpenAI lists API token prices per one million tokens as follows:
| Model | Input | Output | Model ID |
|---|---|---|---|
| GPT-6 Sol | $2.00 | $10.00 | gpt-6-sol |
| GPT-6 Luna | $0.10 | $0.50 | gpt-6-luna |
The announcement characterizes both models as 50% cheaper than the corresponding GPT-5.6 promotional pricing. Its comparison table moves Sol from $4 input and $20 output to $2 and $10. For Luna, it lists a move from $0.20 input and $1.20 output to $0.10 and $0.50. These are API token prices, not a promise about how many Codex tasks a subscription includes. Tool calls, request size, reasoning effort and the amount of generated output can all affect the cost of completing real work.
The price difference between Sol and Luna is large enough that routing every request to Sol would be difficult to justify for many production systems. At the same time, a cheap token does not guarantee a cheap completed task if a weaker fit causes retries, extra review or failed actions. Our GPT-6 pricing guide explains the broader distinction between token pricing and product access. This News release does not change the site’s calculator or evergreen rate verification automatically.
GPT-6 Sol vs Luna: a practical selection rule
OpenAI publishes benchmark results across professional workflows, coding, factuality and computer use, but those results come from specific evaluation setups and effort levels. They should not be treated as a guarantee for a team’s own repository or business process. A better first decision is to match the model to the consequence of failure and the amount of ambiguity in the task.
Use Sol first when the job involves unclear requirements, multiple tools, a long chain of dependent decisions, or a code change that must survive review. Examples include planning a migration, tracing a production bug across services, reconciling conflicting documents, or coordinating an agent workflow with several handoffs. Sol’s role is not merely to produce longer answers; it is the higher-capability option when the system must decide what to do as well as execute it.
Use Luna first when the task is constrained, repeated and easy to check. Examples include classifying a known set of records, applying a fixed formatting transformation, running a documented test matrix, extracting fields from consistent inputs, or producing first-pass summaries for later review. The low API price makes Luna attractive for volume, but the workflow should still define a stopping rule and a review sample.
This is our editorial selection method, not an OpenAI guarantee. Test both models on the same representative cases and compare completion quality, review time and total cost. If Luna needs frequent escalation, route only the ambiguous cases to Sol. If Sol adds no measurable value to a stable task, keep the cheaper route. Readers building an API test can use our GPT-6 API overview as a separate implementation starting point.
Where GPT-6 Sol and Luna are available
OpenAI says GPT-6 Sol and GPT-6 Luna are available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users. Free and Go users can access GPT-6 Luna in the desktop app. At launch, the models are not available in the ordinary Chat surface. API developers can request them with the two model IDs listed above.
The rollout is gradual, so an eligible account may not show the models immediately. OpenAI’s stated instruction is to try again later if they have not appeared in ChatGPT Work or Codex. That makes a missing selector during launch day an availability observation, not proof that an account is permanently excluded. Our GPT-6 access guide covers the product surfaces separately from this dated rollout.
For a team introducing either model, record the surface, plan, model label and date before comparing results. A Codex subscription run and an API request use different billing and access paths, even when they expose similarly named models. Do not infer API entitlement from seeing a model in the desktop selector, or subscription inclusion from a successful API request.
What the launch does not establish
OpenAI’s benchmark claims are vendor-reported results. The announcement notes that research or API evaluations can differ from production ChatGPT because system prompts and available tools differ. It also says the difficult alignment evaluations described on the page are not measurements of typical failure rates. Those caveats matter when translating a chart into a purchasing or deployment decision.
The announcement does not state that GPT-5.6 is retired, set a universal migration deadline, or prove that every existing prompt should move unchanged. Teams should keep a saved baseline, test a small group of representative tasks, and inspect factual errors, tool behavior and review burden separately. A model can be cheaper per token while changing the shape of its answers or requiring different checkpoints.
The confirmed update is narrower and more useful than the rumor: GPT-6 now has two additional official operating tiers, published API prices, concrete model IDs and stated product availability. Sol is the higher-capability daily-work option in our editorial routing method; Luna is the lower-priced API option. Astra remains OpenAI’s top choice for the hardest work. The next step is not to replace every route at once, but to measure where each tier earns its place.