Short answer
Choose GPT-6 Sol for demanding work such as coding, multi-step agent tasks and professional documents, and GPT-6 Luna for simple, repeated jobs at volume, such as sorting, extracting and summarizing. Luna costs about a twentieth of Sol per token, so test whether it is good enough for your task before paying for Sol.
Follow-up question How do they differ in cost and speed?
A week ago, on September 22, OpenAI added two models to its GPT-6 family. GPT-6 Astra, released on September 3, remains the top model. GPT-6 Sol is the strong middle option, and GPT-6 Luna is the light, inexpensive one. OpenAI says it trained both with methods similar to Astra's, so they carry over much of its progress on facts, coding and careful behavior at a lower price.
OpenAI's developer guide sums up the choice in a sentence each: Astra for the highest capability, Sol for strong reasoning on demanding tasks, and Luna for efficient, repeatable work at scale. The rest of this guide turns that into practical decisions.
Where you can use them
Anyone on a paid ChatGPT plan (Plus, Pro, Business, Enterprise or Edu) can pick either model inside ChatGPT Work and inside Codex. People on the Free or Go plans get Luna, but only in the ChatGPT desktop app. At launch, OpenAI said neither model was yet available in the regular Chat view. Developers can call them in the API as gpt-6-sol and gpt-6-luna.
A task-by-task guide
| Task | Pick | Why |
|---|---|---|
| Fixing bugs or building features with Codex | Sol | Coding agents run long, multi-step jobs where mistakes compound |
| Drafting a report, proposal or analysis | Sol | Professional writing benefits from stronger reasoning |
| An agent that works across several apps | Sol | OpenAI highlights Sol's results on business workflows |
| Sorting support tickets into categories | Luna | A fixed choice, repeated thousands of times |
| Pulling names, dates and totals from documents | Luna | Simple extraction at volume |
| Summarizing meeting notes or emails | Luna, then check | Usually enough; move to Sol if summaries miss points |
| A quick factual question you'll act on | Sol | Fewer factual errors, by OpenAI's measure |
| The hardest research or computer-use tasks | Astra | OpenAI still calls Astra its best model overall |
The table reflects OpenAI's own descriptions and benchmark claims, which outside testers have yet to confirm. Treat it as a starting point, not a verdict.
On factual accuracy, OpenAI reports that Sol makes roughly half as many mistakes as GPT-5.6 Sol on a test built from conversations where users had flagged errors, and that Luna, at higher effort settings, matches GPT-5.6 Sol. On a software-engineering benchmark called DeepSWE, it reports 68.8% for Sol and 66.6% for Luna, both at maximum effort. Those figures suggest Luna is a capable model, not a toy, which is why testing it first is worthwhile.
Try it: Pick one task you do often. Run the same five examples through Luna and through Sol, and score each answer as usable, needs editing or wrong. If Luna's scores match Sol's, you have found a job where you can save money.
A worked cost example
Prices are per million tokens, the small chunks of text (roughly three-quarters of a word each in English) that models read and write.
| Model | Input, per million tokens | Output, per million tokens |
|---|---|---|
| GPT-6 Sol | $2.00 | $10.00 |
| GPT-6 Luna | $0.10 | $0.50 |
Suppose a small online shop wants to summarize 1,000 customer emails, each about 2,000 tokens long, into 300-token summaries. The figures are illustrative:
- Luna: 2 million input tokens × $0.10 = $0.20, plus 300,000 output tokens × $0.50 per million = $0.15. Total about $0.35.
- Sol: 2 million × $2.00 = $4.00, plus 300,000 × $10.00 per million = $3.00. Total about $7.00.
Both are small sums at this scale, but the ratio holds as volume grows: for the same job, Luna costs about one-twentieth as much. Higher reasoning effort makes the model think longer and write more tokens, so real costs rise with the effort setting you choose.
How do they differ in cost and speed?
Cost. Per token, Luna's prices are one-twentieth of Sol's for both input and output. Both models are half the price of their GPT-5.6 equivalents at GPT-5.6's promotional rates, a cut OpenAI attributes to more efficient caching and inference. For developers, prompt caching helps too: when requests repeat the same opening text, such as long instructions, the repeated part is billed at a 90% discount, and OpenAI says GPT-6 caches more of it by default. Agents that resend the same context on every step benefit most.
Speed. OpenAI presents both models as faster and more affordable than Astra, but its announcement gives no speed figures for either one. What you control is the reasoning effort, the setting for how long a model thinks before answering. Sol and Luna support an effort of none, which Astra does not, so they can answer straight away when a task needs no deliberation. In general, a smaller model at low effort answers fastest, which makes Luna the natural fit for live features such as classifying a message while a customer waits.
The practical rule: start with Luna at low effort, measure quality on real examples, and move up to Sol, or raise the effort, only where the results fall short. Keep Astra for the jobs where neither is good enough.
Sources
- Introducing GPT-6 Sol and Luna — OpenAI, September 22, 2026
- Model guidance: Using GPT-6 — OpenAI (API documentation), September 26, 2026
- Announcing GPT-6 Sol and GPT-6 Luna in the API, Codex and ChatGPT — OpenAI Developer Community, September 22, 2026