GPT-5.6 explained: how Sol, Terra and Luna differ
Sol, Terra and Luna are versions of the same AI family. Learn how their intended uses and costs differ, and why they are separate from ChatGPT subscriptions.
By Monolith
GPT-5.6 has three versions: Sol, Terra and Luna. Think of them as choices a developer can make when connecting AI to a business process. OpenAI positions Sol for difficult professional work, Terra for a balance of capability and cost, and Luna for large volumes of simpler tasks. They are not three ChatGPT subscriptions you need to buy.342
Match the model to the step in the job
Suppose your agency receives a batch of customer feedback. One task is to sort each comment into a category. Another is to explain why customers are unhappy and recommend changes. The second task calls for more interpretation than the first.
A developer could try a lower-cost model for sorting and a more capable option for the analysis. You do not have to manage that switch manually each time; software can make the selection according to agreed rules. The business decision is whether the extra capability improves a result enough to justify its cost.
Editorial correction, September 11: the previous headline treated these variations as purchasable business tiers. This article now focuses on the model release and its practical implications. The original publication date is unchanged.
Release timing and what changed
OpenAI previewed the family on June 26 and announced wider availability for July 9. The release brought more capable tool-based work into a family with several price points. It also made the naming more explicit: Sol roughly fills the former flagship role, Terra the mini role and Luna the nano role.134
A developer can connect these models to a website or internal tool through an API, a software connection for sending instructions and receiving a result. All three accept text and images and reply with text. You can supply a document or screenshot, but creating a finished presentation or changing a website also requires the appropriate tools.342
| Detail | What it means |
|---|---|
| Default API name | gpt-5.6 selects Sol. |
| Context window | 1,050,000 tokens: the combined working space for information in a request. |
| Output limit | Up to 128,000 tokens for generated output. |
| Long requests | Above 272,000 input tokens, higher rates apply to the whole request. |
What the variations are for
| Variant | Documented position | Candidate workload to test |
|---|---|---|
| Sol | Flagship, roughly earlier unsuffixed tier | Complex professional work with explicit review criteria |
| Terra | Balance of intelligence and cost, roughly mini | Routine analysis requiring more than extraction |
| Luna | Cost-sensitive volume, roughly nano | Classification and structured extraction with validation |
Imagine a process that sorts customer messages and drafts answers to unusual questions. Sorting may need a less expensive model than a difficult answer. Choosing a model for each step is called routing. Test whether the less expensive option gets that step right before adopting it; the lowest usage price does not always produce the lowest total cost.
Build in a way to handle uncertainty. If a message does not fit the categories, the system can flag it for a person instead of forcing a choice. Those unusual examples are especially useful in a trial because they show where the cheaper option needs help. A process that works only on the easiest messages may create more manual work than expected.
Current API access and costs
| Variant | Input | Cached input | Output |
|---|---|---|---|
| Sol | $4 | $0.40 | $20 |
| Terra | $2 | $0.20 | $12 |
| Luna | $0.20 | $0.02 | $1.20 |
Sending more documents can make a request more expensive in two ways: there is more material to process, and a sufficiently large request can move into a higher pricing bracket. Include the documents needed for the task instead of attaching an entire company archive by default.
| Detail | What it means |
|---|---|
| Sol promotion | The published promotional rate is available at least through November 21, 2026. |
| Above 272K input tokens | Twice the input rate and 1.5 times the output rate, applied to the full request. |
| Cache writes | Storing eligible input for reuse costs 1.25 times the uncached input rate. |
A model choice, a ChatGPT subscription and an API account allowance are different decisions. The model determines which system handles the work. The subscription or account rules determine how you access it and how much you can use. The documentation also says fine-tuning, additional training to adapt the model, is not supported.3
A practical routing decision
Choose one task and examples with answers you can check. You could try Luna for sorting inquiries into categories, Terra for combining information into a summary and Sol for a difficult coding task. These are trial ideas, not results of Monolith testing. Count unsuccessful attempts and review time alongside usage fees to see which option produces useful work at a reasonable cost.
Measure useful results, not only the fee per request. If a low-cost draft takes several minutes to correct and another version needs only a quick check, the second may be the better business choice. For a simple sorting task, the less expensive option may already meet your requirements. Let the examples decide.
Keep the same review rules whichever model you choose. A more capable model can still send an incorrect reply or change the wrong record if you give it permission. GPT-6 Astra is a separate release; its arrival does not turn Sol, Terra and Luna into subscription plans.
Terms used in this article
| Term | Plain-language meaning |
|---|---|
| Token | A small unit of information counted by the model. Usage fees often depend on how many are read and generated. |
| Input / output | The information you send / the response the model generates. |
| Cached input | Eligible information stored for reuse, sometimes charged at a lower rate. |
| Benchmark | A defined test. Its score depends on the tasks, settings and supporting tools. |
| Context window | How much information fits into one request; it is not a guarantee of perfect recall. |
Are Sol, Terra and Luna different ChatGPT plans?
What changes above 272K input tokens?
Read for this feature. The numbers match the markers in the text.
- OpenAI GPT-5.6 preview announcementcommunity.openai.com
- OpenAI GPT-5.6 Luna model documentationdevelopers.openai.com
- OpenAI GPT-5.6 Sol model documentationdevelopers.openai.com
- OpenAI GPT-5.6 Terra model documentationdevelopers.openai.com
Want to apply this to a project? Here is the related service.
25 AI instructions to try on everyday work
Prompts are instructions you give an AI tool. This PDF includes 25 examples for inquiries, marketing and admin. Adapt them to your task and check the results.