The short version
● Astra $8 in, $40 out per million. For complex coding and professional work.
● Sol $2 in, $10 out per million. For text, images and tool use.
● Gap Astra costs four times as much on both input and output.
● Rule Sol until it visibly fails. Astra only where the failure costs more than 4x.
OpenAI shipped both on 18 November and the naming does not tell you which to pick. The price does. Astra is four times Sol on input and four times on output, which means the question is never "which is better" but "is this task worth four times the bill".
The pricing, per million tokens
| GPT-6.2 Astra | GPT-6.2 Sol |
|---|
| Input | $8.00 | $2.00 |
| Cached input | $0.40 | $0.10 |
| Output | $40.00 | $10.00 |
| Model ID | gpt-6.2-astra | gpt-6.2-sol |
| Built for | Complex coding, professional work | Text, images and tools |
| Image input | Yes | Yes |
| Batch and Flex | Yes | Yes |
Both run through Responses or Chat Completions, both accept text and image input and return text, and both support Batch and Flex. The capability surface is the same. The difference is depth of reasoning against cost.
Astra got cheaper, which is the part people missed
Astra's input and output are 20% below the previous GPT-6 Astra, and cached input is 60% off. If you benchmarked Astra on the older pricing and ruled it out, the arithmetic has changed. Sol holds the GPT-6.1 Sol rates unchanged.
How to decide without guessing
Run the task on Sol first and look at where it actually breaks. Most production work — summarising, extraction, classification, routine code, tool calling — does not improve enough on Astra to justify a 4x bill. The places that do are long multi-step reasoning chains, hard debugging where a wrong answer costs an engineer an afternoon, and professional work where being wrong is expensive in a way that is easy to price.
A practical pattern: route by default to Sol, and escalate to Astra on a retry when the first answer fails a check you can run cheaply. You pay the 4x only on the fraction that needs it.
The specialist variant
OpenAI also ships Astra for Law, model ID gpt-6.2-astra-law, built on GPT-6.2 Astra. On Harvey's LAB-AA benchmark it scores 20.8% full credit against base Astra's 15.0%, and it answers 112 of 208 questions on Vals AI's Legal Research Bench. Access runs through a request process for ChatGPT Enterprise customers and via the API.
Who can actually use them
In ChatGPT, Astra and Sol are in Work and Codex. Pro 500 gets Ultrafast access with either model, and Pro 200 gets Ultrafast for the first time with Sol only. Enterprise and Edu workspaces get an admin opt-in preview in ChatGPT Work and Codex desktop, off by default, with US and EU inference supported — it is not available in workspaces that require UAE inference, and enabling it does not add either model to regular Chat.
FAQ
What is the difference between GPT-6.2 Astra and Sol?
Astra is built for complex coding and professional work at $8 input and $40 output per million tokens. Sol is built for text, images and tool use at $2 and $10.
Is GPT-6.2 Astra worth four times the price?
Only for long reasoning chains, hard debugging and professional work where a wrong answer is expensive. For routine production work, Sol is usually enough.
Did GPT-6.2 Astra get cheaper?
Yes. Input and output are 20% below GPT-6 Astra and cached input is 60% off. Sol keeps the GPT-6.1 Sol rates.
Do both models accept images?
Yes. Both take text and image input, return text, and support Batch and Flex through Responses or Chat Completions.
What is gpt-6.2-astra-law?
A legal variant built on GPT-6.2 Astra. It scores 20.8% full credit on Harvey LAB-AA against base Astra's 15.0%, and answers 112 of 208 Vals AI Legal Research Bench questions.