이 아티클은 아직 번역되지 않아 영어로 표시됩니다.
GPT-6.1 Sol vs GPT-6 Astra: Near-Astra Results for One Fifth of the Price?
GPT-6.1 Sol costs $2 in and $10 out per million tokens, a fifth of GPT-6 Astra. OpenAI's numbers for coding, office work and science, plus 3 catches.
GPT-6 Astra costs $50 per million output tokens. The new GPT-6.1 Sol costs $10. That is five times cheaper, and on one hard coding test Sol matches Astra. So is Astra a waste of money?
Not quite. Every number here comes from OpenAI's own announcement at DevDay on September 29, 2026. This is not a hands-on test.
▶ WATCH ON YOUTUBEWatch the video
▶ WATCH ON YOUTUBEFull video
The lineup and prices
| Model | Role | Input | Output | Cached input |
|---|---|---|---|---|
| GPT-6 Astra | Flagship, smartest | $10 | $50 | $1.00 |
| GPT-6.1 Sol | The middle | $2 | $10 | $0.10 |
| GPT-6 Luna | Small, fast, high volume | $0.10 | $0.50 | $0.01 |
Prices are per million tokens.
Why cached input matters: cached input is text the model has already seen, such as a long system prompt or a whole codebase. Agents send that context again on every step. For agents, the $0.10 price matters a lot.
What OpenAI's benchmarks say
| Test | What it measures | Sol's result |
|---|---|---|
| DeepSWE v1.1 | Long software tasks in real codebases | Matches Astra at about 1/5 the cost per task. +6.4 points over the previous Sol |
| GDP.pdf | Questions about messy professional documents | Beats Opus 5.5 at under half the cost. Near Astra at about 1/5 the price |
| AutomationBench | Business workflows with 47 tools | +2.2 points over Opus 5.5 at about 1/3 the cost |
| OSWorld 2.0 | Using a computer like a person | 2.1 points behind Astra at about 1/7 the cost |
| Terminal-Bench Science | Real research work | $5.47 per task vs $23.80 for Astra. Astra still scores highest, 68.1% |
Even OpenAI says to use Astra for the hardest science.
Does it make things up?
- On tricky chats where an older model got facts wrong, errors dropped from 11.4% to 7.7% at low effort.
- OpenAI broke the agent's search tool on purpose to see if the model would admit it or guess. How often each failed to mention the broken tool:
| Model | Failed to mention it |
|---|---|
| GPT-6 Astra | 1.5% |
| GPT-6.1 Sol | 2.1% |
| Previous Sol | 4.9% |
| GPT-6 Luna | 28.7% |
These are deliberately hard tests. They show Sol behaves much more like Astra than like the budget model.
The three catches
- Not in regular ChatGPT chat yet. It is in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu, and in the API as
gpt-6.1-sol. - Every benchmark is OpenAI's own. Wait for independent tests.
- "Matches Astra" is one coding test. Everywhere else it is close, not equal.
Which one to pay for
| Your job | Use |
|---|---|
| Coding agents | Sol |
| Documents and business automation | Sol |
| Agents that re-read huge context | Sol, for the cache price |
| The hardest research and science | Astra |
| Simple, high-volume jobs | Luna |
Astra is still the top model. For most people, Sol is the one you will pay for.
Links
- Full video: https://youtu.be/Mky9VkY27sY
- OpenAI, Introducing GPT-6.1 Sol: https://openai.com/index/introducing-gpt-6-1-sol/
- GPT-6 Astra vs Claude Fable 5.1
- Gemini 4 Argon vs GPT-6 Astra vs Claude: API pricing
Would you switch from Astra to Sol?


