ChatGPT 6.1 Sol explained: what GPT-6.1 Sol is, what it costs, and when to use it

GPT-6.1 Sol is a large language model from OpenAI: a program that reads text (and images) and writes text back. People search for it as ChatGPT 6.1 Sol, because many first meet it inside ChatGPT, but the model's name is GPT-6.1 Sol.

OpenAI's GPT-6 family has three models at three price points. Luna is the cheap, fast one, which OpenAI calls its most efficient model. Astra is the one OpenAI calls its most intelligent. Sol sits in the middle. GPT-6 Sol and GPT-6 Luna came out on 22 September 2026. One week later, on 29 September, OpenAI released GPT-6.1 Sol as an upgrade to Sol. OpenAI's API changelog, its public list of changes, has the entry for that date.

OpenAI's own one-line summary is "Near-Astra performance for complex work at a lower cost." This means: almost as good as its best model on hard tasks, for much less money. Section 3 shows what OpenAI claims, with its numbers labelled as OpenAI's.

When I looked at the misses one by one, I found a problem in my own answer key. My ticket script adds the sentence "This looks like someone is in my account at this moment." to 15 tickets, and 13 of them are about other topics, such as a late parcel or a bill. My prompt describes the account team as "Login, passwords, security, profile, users and permissions, data deletion". So the models often sent those 13 tickets to account (Luna 6 times, Sol 9 times), which is a fair choice, and my answer key called it wrong.

All 6 of Luna's misses and 9 of Sol's 10 were on those 15 tickets. On the other 285 tickets, Luna matched the key on all 285 and Sol on 284. Counting ticket by ticket, Luna alone was right on 5 tickets and Sol alone on 1. A standard test for paired results like this (McNemar's test) gives p = 0.22, which means a gap this size would often happen by chance. So the fair reading is: both routed equally well, and Sol cost about 21 times as much and took about twice as long.

My numbers do not prove that Sol is a worse model, or no better. It is one task, on one date, with made-up tickets and one prompt. Sol ran at low effort. OpenAI's claims are about long, hard tasks such as coding, which I did not test.

A second test I ran on 5 October points the other way. On a short job of spotting duplicate company records, I tried two wordings of the prompt on fresh data. With my original wording, Sol found 100% of the real duplicates and Luna 98.9%. With a rewritten wording, Sol found 99.8% and Luna only 92.2%. So Luna was more sensitive to how the prompt was worded, and Sol was steadier.

Nor can my numbers tell you what your bill will be. My tickets averaged 363 input tokens. A long document, a big chat history or a high reasoning effort changes the cost a lot. The next section does the math with the numbers in the open, so you can redo it with yours.

Release date, prices, context window, reasoning settings and where Sol is available come from OpenAI's launch post, API model pages, pricing page and API changelog, all read on 5 October 2026 and linked below. The benchmark results are OpenAI's and I did not rerun them.

The 300-ticket numbers are from my own run on 4 October 2026, read again from the results file for this page. The answer-key problem, the reasoning tokens and the cascade replay come from the same run's per-ticket records. The single API call is from 5 October 2026 and was recorded. I did not test Sol inside ChatGPT, at higher reasoning effort, on long documents, or on coding tasks.