Models
GPT Models in 2026: Astra, Sol, or Luna — Which Should You Choose?
GPT is OpenAI’s model family, which powers ChatGPT. In fall 2026, it has three tiers: flagship GPT-6 Astra, general-purpose GPT-6 Sol, and lower-cost GPT-6 Luna. The difference is not just a version number; it is the balance of capability, speed, and cost per task. This guide explains what each model can do, what it costs, where it may refuse, and who it suits.
Which GPT models are current?
OpenAI introduced GPT-6 Astra in September 2026 as its most capable model. The release was delayed: after an incident involving Hugging Face in July, OpenAI added safeguards and initially showed the model to a small group of organizations. Sol and Luna launched on September 22 and filled out the lower tiers, leaving Astra at the top.
OpenAI describes Sol and Luna as lasting capability tiers rather than one-off versions: the number is the generation, and the name is the model class. This makes updates within a class easier to follow without learning a new lineup each time.
| Model | Role | API price per 1M tokens |
|---|---|---|
| GPT-6 Astra | Flagship for complex reasoning, code, and scientific tasks | Contact for pricing; limited initial access |
| GPT-6 Sol | Workhorse for code and knowledge tasks | $5 input / $30 output (cached input is 90% cheaper) |
| GPT-6 Luna | Fast and low-cost for high-volume work | $0.10 input / $0.50 output after the price cut |
How capable are they?
Astra sets OpenAI records in coding, tool use, and abstract reasoning: on the company’s internal “expert work” benchmark it scored 64.6%, compared with 52.6% for Claude Fable 5.1, while its estimated API cost was about one-third lower. Sol is also competitive: it beat Claude Fable 5 on Artificial Analysis’s agent index while using less than half as many output tokens.
The practical takeaway: on long tasks such as codebase refactoring, analyzing hundreds of pages, or multi-step research, Sol and Astra can save tokens by using fewer words on the way to an answer. Luna is less suited to that level of work, but is more than enough for routine drafts, translation, and simple coding questions.
Safety and refusals
OpenAI says Astra is its most aligned model: it is less likely to give misleading answers or overreach on impossible tasks. The trade-off is tighter cybersecurity safeguards. Because of its strong offensive capabilities, the launch included warnings and staged access. The model may politely refuse some gray-area requests in chat; that is model behavior, not an interface bug.
Refusals on everyday questions are less common than before: OpenAI has worked to reduce false refusals and sycophancy. But if wording makes a task sound dangerous, the model may ask what you mean or refuse even when your intent is harmless.
What does a task cost?
Sol’s API price is $5 per million input tokens and $30 per million output tokens. Cached input costs about one-tenth of the input price, and OpenAI guarantees a cache lifetime of at least 30 minutes. This matters for chatbots and agents: a large, unchanged context of instructions or reference material can be reused at very low cost.
After its price cut, Luna costs $0.10/$0.50 per million tokens. OpenAI says its quality on some tasks is comparable to flagship models from a year earlier. For bulk work such as price lists, reviews, and classification, it can offer strong value per result. ChatGPT Plus and Pro subscriptions use message limits rather than token prices; OpenAI balances model load behind the scenes.
How the models work
- API context windows of up to 1 million tokens can fit a codebase or book in one request.
- Prompt caching with explicit breakpoints offers a 90% discount on cached reads for long conversations.
- Programmatic Tool Calling in the Responses API lets the model write and run coordinator code for tools, filtering intermediate data.
- In ChatGPT, a model router decides when to spend longer reasoning; you can also ask it to think carefully or turn on Thinking mode.
- Limit: availability through aggregators changes, and not every provider offers the latest models right away.
Who should use each model?
If you are not a developer and want a capable everyday chat, ChatGPT’s free tier and Luna or Sol are often enough. Sol is a sensible default for programmers and analysts; Astra makes sense when the task is genuinely difficult and the cost of an error is high.
The fairest way to compare flagship models is to use the same prompts. Try a familiar question in the Any LLM model catalog; it usually lists several current OpenAI models, and you can switch between them in the chat. For GPT’s main competitor, see our separate Claude model guide.
Frequently asked questions
Are GPT-6 and ChatGPT the same thing?
No. ChatGPT is a product, and GPT models run inside it. OpenAI gradually adds different models from its lineup to the product.
Which GPT model is the most capable?
GPT-6 Astra. For most everyday tasks, however, the difference from Sol may be hard to notice while the price difference is clear.
Why does GPT refuse some questions?
OpenAI safety rules apply partly at the model level and partly at the service level. They cannot be removed entirely, but a clearer restatement of a harmless request may help.
Can I use GPT for free?
Yes. ChatGPT has a free tier with limits; regular use may require a subscription or API key.
How is GPT different from Gemini?
It depends on the task. GPT has traditionally been strong at coding and following instructions, while Gemini focuses on multimodality and large-context pricing. Compare both on your own prompt; our Gemini guide covers its strengths.