"GPT-6" posts on X went from about 1,660 a day on Oct 18 to nearly 2,500 on Oct 21. A lot of them are asking the same thing: which GPT-6 am I actually using? OpenAI shipped it as a family with three names, rolled out over several weeks, so the confusion is fair.
The timeline
OpenAI's own pages give this order:
- GPT-6 Astra launched first. It went to a limited set of organizations, then to paid ChatGPT plans and the API.
- Sept 22, 2026: OpenAI added GPT-6 Sol and GPT-6 Luna to the family.
- Sept 29, 2026: GPT-6.1 Sol arrived.
- Oct 7, 2026: GPT-6 with Intelligent UI started rolling out in ChatGPT's Chat tab for Plus, Pro, Business and Enterprise. Free and Go started the next day.
Which model is which
| Name | Where you meet it | Who gets it |
|---|---|---|
| Astra | API (gpt-6-astra), Azure, Bedrock; also ChatGPT paid plans |
Developers; Plus, Pro, Business, Enterprise |
| Sol | ChatGPT Chat tab | Plus, Pro, Business, Enterprise |
| Luna | ChatGPT Chat tab | Free and Go |
If you pay for ChatGPT, everyday chat runs on Sol. If you're on Free or Go, it runs on Luna. Astra is the heavy model for hard work, and it's the one you call from code. OpenAI says ChatGPT has more than 1.2 billion weekly users, so Luna is the GPT-6 most people will actually meet.
What "Intelligent UI" is
OpenAI trained GPT-6 to answer with more than text. A reply can include graphics, tappable buttons, forms, charts and small interactive tools. A comparison can show up side by side. A savings question can come back as a calculator you can play with.
Under the hood, OpenAI built a library of streamable components and a compiler that renders the interface while the model is still generating. When plain text is the best answer, ChatGPT can still just give you text.
The other speed claim is narrower than it sounds. For questions that need web search, OpenAI says GPT-6 Instant starts answering 44% sooner on average than GPT-5.6 Instant. That's about when the answer starts, not how long the whole task takes.
Astra pricing for builders
From OpenAI's Astra page:
- Model ID:
gpt-6-astra(also on Microsoft Azure and Amazon Bedrock) - Input: $10 per million tokens
- Output: $50 per million tokens
- Cache reads and writes: billed at separate rates
- Fast mode: up to 2x the speed at 2x the Standard price
That puts Astra at the expensive end of frontier APIs. Agents produce a lot of output tokens, and output is the $50 side of the bill. Compare cost per finished task, not list price. Our LLM API pricing comparison and prompt caching guide walk through the math.
The benchmarks are Astra's
These are the numbers people quote as "GPT-6 scores". They're from the Astra announcement:
| Benchmark | What it tests | Astra | GPT-5.6 Sol |
|---|---|---|---|
| OSWorld 2.0 | Using a desktop computer | 72.6% | 65.7% |
| Terminal-Bench 4.0 | Terminal tasks | 57.9% | 37.3% |
| DeepSWE v1.1 | Software engineering | 74.1% | 72.7% |
OpenAI ran these itself and says scores are "the maximum at any effort". It also notes that production ChatGPT can differ because of system prompts and tools. The scores tell you what Astra can do at full effort. They aren't a measurement of the Sol or Luna in your chat window.
What the ChatGPT update didn't change
The Oct 7 post says the update applies to the Chat experience only. The models powering Work and Codex were not changing in that release. If you use Codex, check the model in your session rather than assuming. Our Codex CLI guide shows how with /status.
Should you switch?
In ChatGPT you don't have to do anything. Your plan picks the model.
On the API, treat Astra like any new model. Run your own evals on real tasks, watch the output-token bill, and route simple steps to something cheaper. Our guide to picking a coding agent covers how to stay portable so you can swap models without a rewrite.
