Hey everyone! OpenAI released GPT-6.2 Sol today, and Ultra AI in AtomAI is already running on it. If you use Ultra AI or Research, your next question is answered by GPT-6.2 Sol. There is nothing to switch on and nothing extra to pay.
OpenAI calls it "an efficient and versatile model for everyday work". Only Ultra AI has moved. Everyday questions on every plan still run on GPT-6 Luna.
What changed from GPT-6.1 Sol
The biggest gains are on long, practical tasks: the kind of multi-step work where a model has to plan, check itself and finish the job. Here is how GPT-6.2 Sol moved against the model it replaces.
| Benchmark | What it measures | Change from GPT-6.1 Sol |
|---|---|---|
| GDPval v2 | Real work tasks from professional occupations | +5.3 points |
| Terminal-Bench-Science | Scientific workflows run in a terminal | About 57% to about 68% |
| Artificial Analysis Intelligence Index | An average of nine independent evaluations | +2 points, to 57 |
| Artificial Analysis Coding Agent Index | Coding tasks completed as an agent | -3 points |
Source: OpenAI and Artificial Analysis, as reported by The Decoder
GDPval and Terminal-Bench-Science are from OpenAI. The Terminal-Bench-Science figures are read off its launch chart, so treat them as approximate. The two indexes are run independently by Artificial Analysis.
Not every number went up
On the Artificial Analysis Coding Agent Index, GPT-6.2 Sol scored three points lower than GPT-6.1 Sol. Most of that came from one test where it left key facts out of its answers more often than its predecessor did. If an Ultra AI answer feels like it skipped a step, ask it to show the full working. It will.
Next to OpenAI's largest models
Astra is OpenAI's largest model family. GPT-6.2 Sol now matches GPT-6 Astra on the independent Intelligence Index, at about 15% of the cost per task, and answers 96% of graduate-level science questions correctly, the same as GPT-6 Astra. That cost is what lets us run it behind Ultra AI without changing your plan.
| Measure | GPT-6.2 Sol | GPT-6 Astra | GPT-6.2 Astra |
|---|---|---|---|
| GPQA Diamond (graduate-level science) | 96.0% | 96.0% | 97.0% |
| Intelligence Index (Artificial Analysis) | 57 | 57 | 62 |
| Vals Index cost per task | $3.18 | $21.28 | $10.68 |
Source: OpenAI GPT-6.2 system card, Artificial Analysis and Vals AI
Same price, same allowance
GPT-6.2 Sol costs exactly what GPT-6.1 Sol did, so nothing about your plan changes. Your Ultra AI allowance goes just as far, and the Faster, Balanced and Deep thinking settings work as before.
What you need to do
Nothing. Turn on Ultra AI and you are on GPT-6.2 Sol. Your chats, files and study sets are where you left them. Free accounts can try it with their free Ultra AI answers.
Log in and give GPT-6.2 Sol your hardest question!
You can read OpenAI's announcement here, and the full benchmark results in the GPT-6.2 system card.




