OpenAI released GPT-6.1 Sol on 29 September, a week after GPT-6 Sol and almost four weeks after GPT-6 Astra. The pitch is Astra-level results at Sol's price. We read OpenAI's numbers, the independent results that came out the same day, and the API documentation for all three models. Here is how they compare.
Prices
GPT-6.1 Sol costs the same as GPT-6 Sol: $2 per million input tokens and $10 per million output tokens. Cached input gets cheaper, from $0.20 to $0.10 per million. GPT-6 Astra costs five times as much, at $10 and $50, with cached input at $1. For GPT-6.1 Sol and Astra, OpenAI's documentation also lists a higher rate for requests over 272,000 input tokens.
The numbers side by side
| Measure | GPT-6 Sol | GPT-6.1 Sol | GPT-6 Astra |
|---|---|---|---|
| Released | 22 Sep 2026 | 29 Sep 2026 | 3 Sep 2026 |
| Input / output, per million tokens | $2 / $10 | $2 / $10 | $10 / $50 |
| Cached input, per million tokens | $0.20 | $0.10 | $1 |
| DeepSWE v1.1 (OpenAI) | 68.8% (max) | 75.2% (high) | 74.8% (high) |
| OSWorld 2.0, offline set, max effort (OpenAI) | 64.4% | 71.4% | 73.5% |
| Intelligence Index, max effort (Artificial Analysis) | 48 | 52 | 53 |
| Cost per index task (Artificial Analysis) | $1.05 | $0.72 | $3.26 |
Two things to keep in mind when reading it. OpenAI ran the DeepSWE and OSWorld tests itself, and the effort settings differ between models on DeepSWE. The last two rows come from Artificial Analysis, which runs its own index of ten tests. The tests OpenAI led with (DeepSWE, OSWorld 2.0 and Terminal-Bench Science) are not part of that index, so the two sources measure different things and still tell a similar story.
What changed
GPT-6.1 Sol is a clear step up from GPT-6 Sol. It gains 6.4 points on DeepSWE, 7 points on OSWorld and 4 points on the Artificial Analysis index. OpenAI also says the share of answers containing a factual error at low effort fell from 11.4% to 7.7%. Artificial Analysis found it uses about 10 to 30% more output tokens than GPT-6 Sol, and its cost per index task still came out lower.
Against Astra it is close. It is level on DeepSWE, 2.1 points behind on OSWorld and one point behind on the index. On Terminal-Bench Science, OpenAI reports 68.1% for both, at $5.47 per task for GPT-6.1 Sol against $23.80 for Astra.
Astra still leads on harder specialist work. In OpenAI's own security tests it scores 85.4% on SEC-Bench Pro against 78.8% for GPT-6.1 Sol, and on a biology troubleshooting test the gap is wider, 63.46% against 47.96%. Artificial Analysis says Astra leads at every effort level on its index. Astra also has a faster tier, Ultrafast, which OpenAI says reaches up to 300 tokens per second at six times the standard price. VentureBeat reports that a GPT-6.1 Sol version of it is due in the coming days.
There is no GPT-6.1 Astra. TechCrunch reports that OpenAI is not releasing one, and cites a Wall Street Journal report that it was held back over safety concerns raised in internal testing.
Who should switch
If you use GPT-6 Sol, move to GPT-6.1 Sol. The price is the same, and the published scores are higher on every test we found. Check one thing first. OpenAI's documentation lists a "none" reasoning setting for GPT-6 Sol, while the GPT-6.1 Sol page starts at "low". If your setup runs with reasoning switched off to keep replies fast, as we did with GPT-6 Luna in our dental booking test, measure response times before you switch.
If you pay for Astra, run your own tasks on GPT-6.1 Sol before you commit to more Astra spend. For coding agents, computer use and document work, the published gap is a few points and the price gap is five times. We would keep Astra for security and science work, and for jobs where the last few points decide whether the output is usable at all.
If you are choosing from scratch, start with GPT-6.1 Sol and move up to Astra only for the tasks it fails. It is the same rule we use with Anthropic's models.
Sources
- OpenAI, Introducing GPT-6.1 Sol
- OpenAI API docs: GPT-6.1 Sol
- OpenAI API docs: GPT-6 Sol
- OpenAI API docs: GPT-6 Astra
- Artificial Analysis, GPT-6.1 Sol replaces GPT-6 Sol after just 7 days
- Artificial Analysis, GPT-6.1 Sol model page
- Vellum, GPT-6.1 Sol benchmarks explained
- DataCamp, GPT-6.1 Sol: features, benchmarks, pricing
- VentureBeat, GPT-6.1 Sol and the Ultrafast tier
- The Next Web, GPT-6.1 Sol at a fifth of Astra's price
- TechCrunch, OpenAI launches GPT-6.1 Sol