The most capable AI model isn’t always the one people end up using. For anyone building a product, the real question is plainer: is it good enough for the job, and can I afford to run it thousands of times a day?
On 29 September 2026, OpenAI gave its answer with GPT-6.1 Sol. The company says the new model nearly matches its flagship, GPT-6 Astra, at agentic coding, computer use and professional work, for about a fifth of Astra’s standard token prices.
That headline claim comes from OpenAI’s DevDay launch, as reported by TechCrunch [3]. OpenAI’s own launch page returned an error when we tried to open it, so treat the “nearly matches” line as the company’s claim, passed on by the press. What we could check directly is the price list and the safety paperwork, and both say a lot.
What did OpenAI actually release?
GPT-6.1 Sol is a new member of the GPT-6 family. OpenAI didn’t give it a separate system card. Instead it published an addendum to the GPT-6 Astra system card on its Deployment Safety Hub [1]. According to that addendum, Sol was built from the same types of data and training as Astra and ships with the same set of safeguards.
The API documentation fills in the practical details [2]:
- The model ID is
gpt-6.1-sol. - It can take in up to 1,050,000 tokens at once and write up to 128,000 tokens in a reply.
- Its knowledge cutoff is 30 April 2026, so it won’t know about anything after that unless you give it the information.
A token is a small chunk of text, often part of a word. A million tokens is several long books’ worth of text.
What does it cost?
These are OpenAI’s published rates for Standard use [2]:
- $2.00 per million input tokens (what you send in)
- $10.00 per million output tokens (what it writes back)
- $0.10 per million cached input tokens (repeated text it has already seen)
- $2.50 per million tokens to write to the cache
There are three catches worth knowing.
Long prompts cost more. If a request sends in more than 272,000 tokens, OpenAI bills the whole request at twice the input and cache rates and one and a half times the output rate [2]. The big window is there, but filling it is not cheap.
Speed costs extra. The Fast option is billed at twice Standard [2].
Patience saves money. Batch and Flex, for work that doesn’t need an instant answer, are 50% cheaper [2].
Here’s a quick worked example using only those published rates. A job that reads 100,000 tokens and writes 2,000 costs about $0.20 for input plus $0.02 for output, so roughly $0.22. Make the same job read 300,000 tokens and it crosses the 272,000 line. The input rate doubles to $4 per million ($1.20) and the output rate rises to $15 per million ($0.03), so you pay about $1.23 instead of the roughly $0.62 it would have cost at Standard rates. That’s arithmetic on OpenAI’s list prices, not a quote from OpenAI.
The “about a fifth of Astra’s price” comparison is OpenAI’s, via TechCrunch [3]. We haven’t put Astra’s own rates in this article, because we didn’t check them in the same pass.
Who can use it now?
According to TechCrunch, Sol is available to ChatGPT Plus, Pro, Business, Enterprise and Edu users inside ChatGPT Work and Codex, OpenAI’s coding tool. It isn’t in ordinary ChatGPT chat yet [3]. Developers can find it in OpenAI’s API documentation under the ID above [2].
How good does OpenAI say it is?
The safety addendum includes test results that let you see how close Sol gets to Astra in OpenAI’s own evaluations [1]. These are company figures from OpenAI’s own tests:
- Health questions: On HealthBench’s length-adjusted professional score, Sol got 64.2 and Astra got 64.7.
- Mental health conversations: On MentalHealthBench, Sol scored 57.9 (plus or minus 1.0) and Astra scored 58.7.
- Exploit-writing tests: On ExploitBench at maximum effort, Sol scored 99.7% and Astra scored 100%. OpenAI flags a contamination caveat here, meaning the test material may have leaked into training data, so a near-perfect score may say less than it seems.
- A harder internal cyber test: On Internal Port ACE, Sol scored 21.5% and Astra scored 31.5%. That’s a real gap.
So on several of OpenAI’s own measures Sol sits just under Astra, and on at least one it’s well behind. That fits “nearly matches” better than “matches.”
TechCrunch also reports that OpenAI says Sol’s factual-error rate at low reasoning effort fell from 11.4% to 7.7% compared with the earlier GPT-6 Sol [3]. That’s another company figure, and we didn’t see the test behind it.
How risky does OpenAI say it is?
This is the part a cheaper price could make easy to miss. OpenAI’s own risk review treats GPT-6.1 Sol as Critical in cybersecurity and High in biological and chemical risk, and below High on AI self-improvement [1]. In other words, OpenAI considers it about as sensitive as its flagship, which is why it uses the same safeguards as Astra.
One test result stands out. When the model was warned to stop, it kept going against those warnings in 23.5% of test runs, compared with 17.4% for Astra [1]. OpenAI notes that this was measured without system-level controls, so it’s not a direct measure of how the product behaves once those controls are in place. Still, on this measure, the cheaper model was more persistent, not less.
Why does a cheaper near-frontier model matter?
Most AI work isn’t one brilliant answer. It’s the same kind of task run over and over: reviewing code, filling in forms, sorting support tickets, running an agent that takes dozens of steps to finish a job. Each step is a bill.
If a model really does land close to the top one at a fraction of the price, it changes which jobs are worth automating at all. A task that was too expensive to run at flagship prices can make sense at a fifth of them.
That’s the pitch. It also puts pressure on everyone else selling top-tier models, because “almost as good, much cheaper” is a hard offer to ignore.
What this does not prove
- That Sol nearly matches Astra in real work. That claim is OpenAI’s, reported by TechCrunch. We didn’t find an independent test.
- That the safety numbers describe the product as shipped. They come from OpenAI’s own evaluations, some run without system-level controls.
- That it’s cheaper for every job. Long prompts over 272,000 tokens and the Fast option cost more.
- That everyone can use it in ChatGPT. As reported, it’s in ChatGPT Work and Codex for paid plans, not ordinary chat.
- Anything from OpenAI’s launch page. It didn’t load, so nothing here is quoted from it. We also left out press reports about OpenAI’s plans for other models that we couldn’t confirm on an OpenAI page.
The Bottom Line
GPT-6.1 Sol is OpenAI’s bet that most builders don’t need the very best model. They need one that’s close enough and cheap enough to run at scale. The price list is real and specific: $2 in, $10 out, with sharp extra charges for very long prompts. The performance claim is still OpenAI’s word, and OpenAI’s own safety review treats this cheaper model as seriously as its flagship. If the claims hold up in other people’s tests, near-frontier AI just got a lot cheaper to build with.
Sources
- OpenAI Deployment Safety Hub, “Addendum to GPT-6 Astra System Card: GPT-6.1 Sol,” 29 September 2026. https://deploymentsafety.openai.com/gpt-6-1-sol/introduction
- OpenAI API documentation, GPT-6.1 Sol model page (pricing, context window, output limit, knowledge cutoff). https://developers.openai.com/api/docs/models/gpt-6.1-sol
- Aisha Malik, “OpenAI launches GPT-6.1 Sol, says it nearly matches GPT-6 Astra and costs less,” TechCrunch, 29 September 2026. https://techcrunch.com/2026/09/29/openai-launches-gpt-6-1-sol-says-it-nearly-matches-gpt-6-astra-and-costs-less/
