Documentation
Which GPT-6 model fits your task
Last updated:
The ChatGPT API serves three current models: GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna. Their quality differs. What that quality costs differs far more, by up to a hundred times. Below is where the line falls and how to pick one for your task without overpaying for nothing.
How Astra, Sol and Luna differ
Our prices, per 1M short-context tokens. OpenAI's official rate is twice that (the struck-through figure in the pricing table is exactly it).
| gpt-6-astra | gpt-6.1-sol | gpt-6-luna | |
|---|---|---|---|
| Best for | The hardest work | Everyday work | High volume |
| Input / 1M | $5.00 | $1.00 | $0.05 |
| Output / 1M | $25.00 | $5.00 | $0.25 |
| Context | 1.05M tokens | 1.05M tokens | 1.05M tokens |
| Max output | 128K tokens | 128K tokens | 128K tokens |
| Knowledge to | April 30, 2026 | April 30, 2026 | May 18, 2026 |
| Released | September 3, 2026 | September 29, 2026 | September 22, 2026 |
When you actually need GPT-6 Astra
OpenAI's most capable model. Also the most expensive one we serve. There is one classic mistake with it: reaching for it 'just in case'. Use it after Sol has already fallen short. On routine work you won't spot the difference, and the bill still grows fivefold.
- You rewrite a module and the change drags dozens of files with it
- You build an agent that decides its own next step
- You read a whole repository at once: the context is needed together, not in slices
gpt-6-astra$5.00 / $25.00 · 1M
GPT-6.1 Sol: why to start here
It trails the flagship by far less than it saves. Start every project here. Quality holds, you saved five times over. It doesn't, you switch with one line in the request: the key and the base URL stay put.
- You ship features, fix bugs, run code review
- You translate and edit copy
- You pull an analytical summary out of a dozen documents
- You answer customers in support
- You run an agent that needs five or six steps
gpt-6.1-sol$1.00 / $5.00 · 1M
GPT-6 Luna wins on volume
A hundred times cheaper than Astra. Its home ground is work spelled out to the comma and repeated by the thousand: no thinking required, just fast and identical. Putting the flagship on that is like chartering a jet to post a letter.
- Sort tickets into categories
- Pull fields out of statements and invoices
- Squeeze a long document into three paragraphs
- Label a dataset that would take a year by hand
gpt-6-luna$0.05 / $0.25 · 1M
The routing rule
Don't guess. Take the cheapest model that clears your quality bar and move up only when it fails. Defaulting to the flagship is the most expensive way to avoid thinking.
- 1Write down what counts as an acceptable answer. Without that there is nothing to compare.
- 2Run the task on gpt-6-luna. If it passes, anything more is overpayment.
- 3If it fails, climb: gpt-6.1-sol first, then gpt-6-astra. Stop at the first one that does the job, and don't go higher out of curiosity.
How to test the difference on your own data
OpenAI's benchmarks say almost nothing about your task. Collect 20–30 of your own examples, awkward ones included, and run them through all three models.
- How many answers went straight into production untouched
- How often you had to re-ask
- What a solved task cost, not what a sent request cost. Those are two different numbers, and the second one always looks prettier
- Response time, if the person on the other end notices it
- What happens on junk: empty input, a fragment, ten pages of text
What the ChatGPT API costs per month
Costed on a typical small-product load: 10,000 requests a month, 2,000 input and 500 output tokens each.
| Model | Per month |
|---|---|
gpt-6-astra | $225.00 |
gpt-6.1-sol | $45.00 |
gpt-6-luna | $2.25 |
A hundredfold gap. That is why you start at the bottom: on half the tasks the expensive model adds nothing but cost.
Older ChatGPT models: GPT-5.6, 5.5 and 5.4
They haven't gone anywhere, they run on the same key, and nobody is switching them off. There is one reason to pick them: a project is already tuned for them and there is no time to migrate.
gpt-5.6-sol- Complex professional work
gpt-5.6-terra- Everyday code and writing
gpt-5.6-luna- Fast, cheap requests
gpt-6-sol- The previous Sol in the GPT-6 line
gpt-5.5- Proven all-rounder
gpt-5.4- Balance of cost and quality
gpt-5.4-mini- Classification and field extraction
Changing the model in a request
The model goes in the `model` field, exactly as in the standard OpenAI API. Your key and base URL stay the same, so moving from one ChatGPT model to another is a one-line change you can make in production.
{
"model": "gpt-6.1-sol",
"messages": [{ "role": "user", "content": "..." }]
}FAQ
Which model should I start with?+
gpt-6.1-sol. It is close to the flagship in quality at a fifth of the price, so it fits most jobs. If the answers fall short, switch to gpt-6-astra with one line.
How does gpt-6-astra differ from gpt-6.1-sol in practice?+
Astra is noticeably steadier when many constraints must be held at once: large refactors, long agent loops, analysing a lot of material. On ordinary development and writing the difference usually doesn't show, and it costs five times more.
Which ChatGPT model is cheapest?+
The cheapest ChatGPT model in our API is gpt-6-luna: on input tokens it costs a hundred times less than Astra. It is built for high volumes of simple requests: classification, field extraction, short summaries.
What does the reasoning setting change?+
It controls how much the model spends on thinking before answering. Higher levels give better results on hard problems and bill more output tokens. On simple tasks a low level saves money with no loss of quality.
Can I use several models in one project?+
Yes, and it is usually the cheapest setup. The key and base URL are shared, so send simple requests to Luna and hard ones to Sol or Astra. Only the model field changes.