Short answer: GPT-6 Luna is the lower-cost first model to test for focused, repeatable work at volume. GPT-6 Sol is the stronger first test for difficult coding and multi-step agent workflows. Neither is automatically the better choice: compare them on your own tasks, and include the cost of corrections in the decision.
OpenAI released GPT-6 Sol and GPT-6 Luna on September 22, 2026. The practical choice depends on three separate questions: where you can use each model, which one meets your quality bar, and what the complete API request costs. The published API token rates do not describe ChatGPT or Codex plan quotas.
GPT-6 Sol vs Luna at a glance
| Question | GPT-6 Sol | GPT-6 Luna |
|---|---|---|
| OpenAI’s stated role | Complex coding and agentic workflows | Focused, high-volume tasks |
| Best first test | Multi-step work where judgment, tool use, or recovery from mistakes matters | Repeatable tasks with a clear acceptance rule, such as extraction, classification, or short summaries |
| Standard API price up to 272K input | $2 input and $10 output per million tokens | $0.10 input and $0.50 output per million tokens |
| Context and maximum output | 1,050,000-token context; up to 128,000 output tokens | 1,050,000-token context; up to 128,000 output tokens |
The task descriptions in this table reflect OpenAI’s product positioning, not a guarantee that one model will pass your evaluation. The official GPT-6 Sol model page describes complex coding and agentic workflows; the GPT-6 Luna model page describes focused, high-volume work.
Both model cards list text and image input with text output, a 1,050,000-token context window, and reasoning effort settings from none through max. They do not list audio or video support. If your application needs either modality, check the current model cards before choosing this pair.
First check where you can use them
OpenAI’s September 22 release note lists GPT-6 Sol and Luna in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users. It says Free and Go users can access Luna in the desktop app. The same note says the models were not yet available in Chat and that ChatGPT availability would roll out gradually. Check the model picker in the product you use rather than assuming that a model visible in one surface is available everywhere. OpenAI’s launch announcement has the current product wording.
For API use, the model IDs are gpt-6-sol and gpt-6-luna. OpenAI’s API changelog lists both for Responses and Chat Completions, with text and image input and text output. If you are choosing an API endpoint as well as a model, check the model page for endpoint-specific tool support; the two decisions are related, but they are not the same.
A ChatGPT or Codex subscription is not billed at the API rates below. Plan access and usage limits belong to the product plan; API token rates apply to API requests. If you are comparing a subscription workflow with an API integration, compare the actual limits and total workflow cost in each product.
GPT-6 Sol and Luna API pricing
The following are OpenAI’s Standard API rates per one million tokens. “Short context” applies to requests with up to 272,000 input tokens. Cache reads and cache writes are billed at different rates, so a single input-price number is not enough for a cached application.
| Model | Context tier | Input | Cached input | Cache writes | Output |
|---|---|---|---|---|---|
| GPT-6 Sol | Up to 272K input tokens | $2.00 | $0.20 | $2.50 | $10.00 |
| GPT-6 Sol | More than 272K input tokens | $4.00 | $0.40 | $5.00 | $15.00 |
| GPT-6 Luna | Up to 272K input tokens | $0.10 | $0.01 | $0.125 | $0.50 |
| GPT-6 Luna | More than 272K input tokens | $0.20 | $0.02 | $0.25 | $0.75 |
USD per 1M tokens. These are Standard rates from the OpenAI API pricing page. The model pages say that once a request exceeds 272K input tokens, the higher input and cache rates and the 1.5× output rate apply to the full request. Batch and Flex are priced at 50% of Standard, and Fast mode at 2x the applicable rate.
That threshold can change a cost estimate materially. For example, 100,000 uncached input tokens plus 5,000 output tokens at the short-context Standard rates costs about $0.25 with Sol ($0.20 input + $0.05 output) or $0.0125 with Luna ($0.01 + $0.0025). This is token arithmetic, not a claim about the cost of a finished task.
For a request with 300,000 uncached input tokens and 10,000 output tokens, the full request uses long-context rates: about $1.35 with Sol ($1.20 input + $0.15 output) or $0.0675 with Luna ($0.06 + $0.0075). The calculation excludes cached tokens, cache writes, service-tier differences, retries, and any separately billed tools. Check the live pricing page before estimating a production workload.
Eligible shared prompt prefixes reused within OpenAI’s 30-minute cache window can receive cached-input rates. Measure cache hits before assuming the discount: a prompt that changes its prefix each run may not see the same savings as one that repeatedly sends a stable prefix. See OpenAI’s GPT-6 prompt-caching guide for the cache behavior and diagnostics.
One launch claim also needs a little arithmetic context. OpenAI describes the new prices as “50% cheaper” than GPT-5.6 promotional pricing. The listed Luna input rate falls from $0.20 to $0.10, exactly 50%; its output rate falls from $1.20 to $0.50, or about 58.3%. Use the actual rate that matches your input/output mix rather than applying one percentage to every part of a bill. For the previous generation’s API rate structure, see the current GPT-5.6 rate table.
When should you test Sol first?
Start with Sol when the task involves several dependent steps, substantial code changes, tool calls, or a costly failure that needs the model to notice and correct its own mistake. Examples include proposing a code change against a real test suite, reconciling conflicting source documents, or completing a workflow that must follow multiple constraints.
That does not mean Sol will always finish in fewer calls or require less review. A complicated workflow can spend more on input context, generated tokens, tools, and retries. Compare the GPT-6 Astra overview and Astra API pricing guide if you also need to decide whether Astra’s higher-capability tier is justified for a subset of work.
When should you test Luna first?
Start with Luna when each item has a clear target output and the same process runs many times: extracting fields from forms, classifying support requests, summarizing routine documents, or generating first-pass labels for human review. Its token rates are substantially lower, which makes it a sensible candidate for a high-volume pilot.
Low token price does not make a wrong result cheap. If Luna causes more corrections, retries, or escalations, the total cost per accepted result may rise. Use a cost per accepted marketing result to compare quality and economics together. For the previous model generation’s task and surface differences, see the previous GPT-5.6 model-selection guide.
A fair Sol-versus-Luna test you can run
- Pick one recurring job. Write down what a correct, usable result must contain and what counts as a failure. Do not begin with a vague “which answer feels smarter?” question.
- Use representative cases. Test a small set of real examples, including edge cases and inputs that usually trigger corrections. A handful of cases is a pilot, not a statistically conclusive benchmark.
- Hold the setup steady. Use the same source material, prompt, tools, output format, reasoning effort, and review rule where the interface permits. Record any setting you cannot keep identical.
- Compare outcomes. Track first-pass acceptance, factual or format errors, human edit time, retries, latency, and billed input/output tokens. If the output is sensitive or consequential, have a person verify it.
- Choose the cheapest model that clears the bar. Route harder cases to Sol only if your results show that the quality gain is worth the difference. Recheck after prompts, tools, prices, or product limits change.
OpenAI publishes benchmark and cost-per-task comparisons, but those are provider-reported results under selected tasks and effort settings. Treat them as a reason to build a test set, not as a substitute for one. Do not compare scores from different effort levels as if they were the same experiment. A useful local question is: “What does one accepted result cost, including review and retries?”
For better controlled prompt comparisons, the GPT-6 Astra prompting guide covers context and instruction hygiene. Keep the test focused on the Sol/Luna choice here; a prompt rewrite can affect results as much as a model switch.
Frequently asked questions
Is GPT-6 Sol better than GPT-6 Luna?
OpenAI positions Sol for complex coding and agentic work, and Luna for focused, high-volume work. That describes intended use, not a universal ranking. Test both on the task and quality bar you care about.
Which model is cheaper?
Luna has the lower API token rates: $0.10 input and $0.50 output per million tokens at Standard short context, compared with Sol at $2 and $10. The 272K threshold changes both models’ rates, and a lower token bill does not guarantee a lower cost per accepted result.
Can I use Sol or Luna in regular ChatGPT?
OpenAI’s September 22 note listed them in ChatGPT Work and Codex, said Free and Go users could access Luna in the desktop app, and said they were not yet in Chat. The note also described a gradual ChatGPT rollout. Check the current model picker because availability can change.
Do Sol and Luna have different context windows?
The official API model pages list the same 1,050,000-token context window and 128,000-token maximum output for both. The context ceiling does not tell you which model will handle your particular long document more accurately; test with representative inputs and keep an eye on the long-context API price threshold.
Sources and related reading
- OpenAI: Introducing GPT-6 Sol and GPT-6 Luna
- OpenAI API changelog
- GPT-6 Sol model details and GPT-6 Luna model details
- OpenAI API pricing
- OpenAI: Better prompt caching for GPT-6
Updated September 23, 2026. This guide summarizes official documentation and does not report an independent DMT benchmark.
