GPT-6.1 Sol Explained: OpenAI’s Near-Flagship Model at One-Fifth of the Price

OpenAI had a strange week in late September. The model everyone expected did not show up, and the model that did show up may matter more for most people anyway. GPT-6.1 Sol arrived at DevDay in San Francisco on 29 September 2026, and it has spent October becoming the model developers actually talk about when cost comes up. Why? Because OpenAI says it gets close to its flagship Astra model on real coding work while costing about a fifth as much.

That claim deserves a careful look. Prices are facts. Benchmarks are the vendor’s own numbers until independent tests confirm them. This post separates the two, so you can judge whether Sol fits your work.

What GPT-6.1 Sol actually is

Sol is an upgrade to GPT-6 Sol, which itself had launched only a week earlier, on 22 September. The 6.1 version is tuned for three kinds of work: writing and fixing code, driving a computer through tasks, and professional work like reading long documents and answering from them. OpenAI’s CEO called it a future daily workhorse, and the pitch is simple: most of the flagship’s ability, at a much lower price.

There is also a backstory worth knowing. OpenAI had lined up a model called GPT-6.1 Astra, but reports say it was pulled the day before DevDay after failing internal safety tests. OpenAI did not ship it. Sol took its place on stage. Treat that as reported rather than confirmed, since OpenAI has not published a full account of it, but it explains why the launch felt sudden.

The price, in plain numbers

API pricing is where Sol makes its case. Per million tokens:

• GPT-6.1 Sol: $2 input, $10 output, $0.10 cached input
• GPT-6 Astra (flagship): $10 input, $50 output, $1.00 cached input
• GPT-6 Luna (lightweight): $0.10 input, $0.50 output, $0.01 cached input

So Sol costs exactly one-fifth of Astra on input and output. The bigger change is cached input, which fell from $0.20 on GPT-6 Sol to $0.10 on 6.1 Sol. That matters because coding agents re-send the same instructions, files, and conversation history on nearly every step. When most of your prompt is repeated text, the cache discount is where the real saving shows up. One pricing analysis put a typical coding-agent step about 16% cheaper on 6.1 Sol than on 6 Sol for that reason.

One catch: very long prompts cost more. Past 272,000 input tokens, the whole request bills at double the input and cache rates and 1.5 times the output rate. If you regularly feed huge codebases in one go, budget for that.

What OpenAI says it can do

OpenAI’s own benchmark results back the “near-Astra” claim on the work that matters to businesses:

• DeepSWE v1.1 (long software-engineering tasks inside real codebases): matches Astra’s score at roughly a fifth of the cost, and beats GPT-6 Sol’s best score by 6.4 points.
• AutomationBench (multi-step business workflows across dozens of tools): 2.2 points above Anthropic’s Claude Opus 5.5 at medium effort, at about a third of the cost.
• OSWorld 2.0 (computer use): within 2.1 points of Astra at maximum effort, at roughly one-seventh of the cost per task.
• Factual accuracy: the error rate on difficult prompts fell from 11.4% on GPT-6 Sol to 7.7% at low effort, a cut of about a third.

Read those numbers with the right label on them: they are OpenAI’s results, not independent tests. They are still useful, because they show where OpenAI thinks Sol is strong — long coding jobs, agent workflows, and computer use — but wait for outside benchmarks before treating any single score as settled.

Where you can actually use it

Sol is not everywhere yet. You can reach it in three places:

• ChatGPT Work on Plus, Pro, Business, Enterprise, or Edu plans.
• Codex, OpenAI’s coding agent, on the same plans — this is where Sol makes the most sense, since coding is its strongest area.
• The OpenAI API as gpt-6.1-sol, with about 1.05 million tokens of context and up to 128,000 tokens of output.

What about the regular ChatGPT window? Not yet. OpenAI says Sol is not available in Chat at the time of writing, and the free API tier does not include it either. If you are on a free plan, the closest option is the lightweight Luna model. And if you use GitHub Copilot, Sol is rolling out to Pro+, Max, Business, and Enterprise tiers.

Should you switch to it?

Ask yourself three questions. First, do you pay for API calls today? If yes, Sol is a straightforward saving over Astra for coding and agent work, and a clear step up from 6 Sol thanks to the cheaper cache. Second, do you build agents that re-send long context? Then the $0.10 cached rate is the feature, not the headline price. Third, do you need the absolute best score on the hardest research tasks? OpenAI itself still points those users at Astra, which keeps the top score on some science benchmarks even at five times the price.

For most developers, the honest answer is that Sol is now the default sensible choice for coding work, with Astra kept for the jobs where a few extra points are worth five times the cost. Try it on a real task from your own codebase before you commit — your own results will tell you more than any benchmark table.