OpenAI logo via Wikimedia Commons
What GPT-6.1 Sol is and where it sits in the GPT-6 family
GPT-6.1 Sol is the newest mid-tier model in the GPT-6 family, and OpenAI announced it on September 29, 2026. In the API it goes by the model name gpt-6.1-sol. The headline is simple: it is priced at one fifth of the flagship while OpenAI says it gets close to that flagship on several of its own tests.
The GPT-6 family has three tiers. GPT-6 Astra is the flagship, the model OpenAI points to for its strongest results, and we covered it in our Astra explainer. Sol is the cost-efficient high-end tier, meant for serious work that does not need the flagship every time. Luna is the fast tier. GPT-6 Sol and Luna both launched in September 2026, and OpenAI cut token prices by half or more compared with the earlier generation. For reference, GPT-5.6 Sol had cost $4 per million input tokens and $20 per million output tokens.
The pricing, line by line
GPT-6.1 Sol costs $2 per million input tokens, $0.10 per million cached input tokens, and $10 per million output tokens, according to OpenAI's API pricing. The cached input rate fell from $0.20, so that line was cut in half.
For comparison, GPT-6 Astra costs $10 per million input tokens, $1 per million cached input tokens, and $50 per million output tokens. Put the two side by side and the ratio is the same on standard input and output: Sol is exactly one fifth of Astra. Cached input follows a slightly different pattern, since $0.10 is one tenth of $1, so Sol's cached rate is even lower relative to Astra's.
A token is a small chunk of text, roughly three quarters of a word in English. Input tokens are what you send to the model, including your instructions and any documents. Output tokens are what the model writes back. Output is priced higher than input on both models, at five times the input rate, which means a task that produces long answers costs more than one that reads a lot and answers briefly.
Why a one-fifth price matters
A price at one fifth of the flagship matters because it changes which jobs are worth running through an AI model at all. A task that costs a dollar on Astra costs twenty cents on Sol, and the gap grows with volume.
Think about a team that runs an AI agent over thousands of support tickets, code reviews, or documents every day. At that scale the model bill is often the largest line item, and a fivefold difference decides whether a project gets approved. Cheaper tokens also let developers retry a failed attempt, run a second pass to check the first, or give an agent more room to work through a problem, because each extra step costs less.
What cached input means and why it helps agents
Cached input is text the model has already processed recently, and when you send the same text again, the provider charges a much lower rate for it. On GPT-6.1 Sol that rate is $0.10 per million tokens, compared with $2 for fresh input.
This matters because many real workloads repeat themselves. An agent that works through a long task usually sends the same instructions, tool descriptions, and earlier steps with every new request. A support bot sends the same company policy document with every customer message. A long chat resends the whole conversation so far each time you add a line. Without caching, you would pay the full input price for that repeated material over and over. With caching, the repeated part costs a twentieth of the normal rate on Sol.
Here is a simple picture. Suppose an agent carries 100,000 tokens of instructions and history into each of 50 steps. That is 5 million input tokens in total. If most of it qualifies as cached, the bill for that portion drops from about $10 at the standard rate to well under a dollar. The exact savings depend on how OpenAI applies caching to your requests, so check the pricing documentation for the rules, such as how long cached text stays available. The general lesson holds: keep the stable parts of your prompt at the start and put the changing parts at the end, so more of each request can be cached.
The Ultrafast tier and what it trades
Ultrafast is a new service tier that delivers up to 300 tokens per second in the API, and it trades money for speed by charging 6 times the standard rate. If you want answers sooner and are willing to pay for it, this is the option. If cost matters more than waiting a few extra seconds, the standard tier is the better fit.
Applying the 6 times multiplier to Sol's standard prices gives a derived Sol Ultrafast cost of $12 per million input tokens, $0.60 per million cached input tokens, and $60 per million output tokens. Those figures come from the multiplier and the standard prices, and you should confirm them on OpenAI's pricing page when the tier goes live.
Availability differs by model. Astra Ultrafast is available now for Pro 500 and Enterprise. Sol Ultrafast is listed as coming soon, so nobody can use it yet. That also means the Sol Ultrafast numbers above are a calculation, and no one has published real-world experience with them.
Who would pay for speed? Interactive products are the obvious case, such as a voice assistant where a delay feels awkward, or a coding tool where a developer is watching the cursor. Batch jobs that run overnight gain almost nothing from it. Notice also that Sol Ultrafast at $12 input and $60 output costs more than standard Astra at $10 and $50. If you need both strong reasoning and speed, you will want to compare Sol Ultrafast against Astra Ultrafast and see which one fits your test cases.
What OpenAI's own evaluations show
OpenAI says GPT-6.1 Sol improves on its predecessor across several tests, but these are OpenAI-run evaluations, and independent testers have not confirmed them. On DeepSWE v1.1, a software engineering test, OpenAI reports that GPT-6.1 Sol beats GPT-6 Sol's best result by 6.4 percentage points. On AutomationBench, which measures how well a model handles automation tasks, it improves by 4.8 points over GPT-6 Sol at medium reasoning. On OSWorld 2.0, a test of operating a computer, it comes within 2.1 points of Astra.
The last result is the one behind the headline question. Coming within 2.1 points of the flagship on that test, at one fifth of the price, is a strong value claim if it holds up. It does not mean Sol matches Astra everywhere. A gap of 2.1 points on one benchmark says nothing about the other tasks you might care about, and OpenAI has not published a head-to-head comparison across every area in the coverage we reviewed.
Where you can use GPT-6.1 Sol today
GPT-6.1 Sol is available in the API and in several ChatGPT products, but it is not yet available in regular Chat. The ChatGPT surfaces that have it are ChatGPT Work, Codex, and the Plus, Pro, Business, Enterprise, and Edu plans.
That gap in regular Chat is worth understanding before you plan around the model. Someone who opens the standard chat window on a free or basic account may not see Sol listed yet, even though a paying developer can call it through the API the same day. Rollouts often reach different products at different times, and OpenAI has not said when regular Chat will get it in the coverage we reviewed.
For context size, the figures we have are for the previous version. GPT-6 Sol has a context window of 1,050,000 tokens and a maximum output of 128,000 tokens. Those are the GPT-6 Sol numbers and are unconfirmed for GPT-6.1 Sol, so check OpenAI's model documentation for the 6.1 limits before you design around them. A context window is the total amount of text the model can consider in one request, and a million tokens is enough for a very large codebase or a stack of long documents.
Sol or Astra: a general way to choose
Choose Sol when cost and volume matter most and your task is well defined, and choose Astra when a wrong answer is expensive or the problem is unusually hard. No published benchmark settles this for your case, so treat this as general guidance.
Sol makes sense for work like drafting and summarizing at scale, routine coding help, structured data extraction, agents that follow a clear process, and anything you run thousands of times. At one fifth of the cost, you can afford to add checks, such as a second pass that reviews the first.
Astra makes sense when you handle fewer, higher-stakes requests: a difficult research question, a tricky bug that has resisted other attempts, a long analysis where small errors compound, or work where a human reviewer costs more than the model. The extra $8 per million input tokens is small if the answer saves hours.
A practical method is to run your own test set through both. Take 30 to 50 real examples from your work, run them on Sol and Astra, and count how often each one gives an answer you would accept. If Sol passes nearly as often, the price difference decides the question. If Astra clearly wins on the cases that matter, pay for it on those cases only, and send the easy ones to Sol. Many teams split traffic this way.
A worked cost example
A job with 1 million input tokens and 200,000 output tokens costs $4.00 on GPT-6.1 Sol and $20.00 on GPT-6 Astra, using standard uncached rates.
Here is the math for Sol. One million input tokens at $2 per million is $2.00. Two hundred thousand output tokens is 0.2 million, and at $10 per million that is $2.00. The total is $4.00.
Now Astra. One million input tokens at $10 per million is $10.00. The 200,000 output tokens at $50 per million come to $10.00. The total is $20.00, which is five times the Sol bill, matching the price ratio.
Scale it up and the gap becomes real money. If you run that same job 1,000 times a month, Sol costs $4,000 and Astra costs $20,000. If part of the input is cached on Sol at $0.10 per million, the Sol total drops further. For the Ultrafast option, the derived Sol figures give $12.00 for the input plus $12.00 for the output, so $24.00 for the same job, which is above standard Astra. Speed has a price, and this is what it looks like in dollars.
Caveats worth keeping in mind
The biggest caveat is that the performance claims come from OpenAI's own evaluations, and independent results may differ. The 6.4, 4.8, and 2.1 point figures describe specific tests under OpenAI's settings, and results can differ from what you see in production ChatGPT or on your own tasks.
Rollout gaps are the second issue. Sol is in the API and several ChatGPT plans, but not in regular Chat, and Sol Ultrafast is still coming soon. If your plan depends on a feature that has not shipped, you are relying on a date nobody has confirmed. Third, prices change. They dropped by half or more between generations and fell again with this release, and the figures in this article reflect the launch announcement and the coverage around it. Confirm current rates before you build a budget, and check whether your account sees the same limits and tiers as the ones described here.
Frequently asked questions
When was GPT-6.1 Sol released?
OpenAI announced GPT-6.1 Sol on September 29, 2026. The API model name is gpt-6.1-sol.
How much does GPT-6.1 Sol cost?
It costs $2 per million input tokens, $0.10 per million cached input tokens, and $10 per million output tokens. GPT-6 Astra costs $10, $1, and $50 for the same three categories, so Sol is one fifth of Astra on standard input and output.
What is the Ultrafast tier?
Ultrafast delivers up to 300 tokens per second in the API at 6 times the standard rate. For Sol that works out to $12 input, $0.60 cached input, and $60 output per million tokens. Astra Ultrafast is available now for Pro 500 and Enterprise, and Sol Ultrafast is coming soon.
Is GPT-6.1 Sol as good as Astra?
OpenAI says it comes within 2.1 points of Astra on OSWorld 2.0, and it reports gains over GPT-6 Sol on DeepSWE v1.1 and AutomationBench. These are OpenAI-run evaluations and not independent tests, so they do not show Sol matching Astra on every task.
Where can I use GPT-6.1 Sol?
It is available in the API and in ChatGPT Work, Codex, and the Plus, Pro, Business, Enterprise, and Edu plans. It is not yet available in regular Chat.
What is the context window?
The previous version, GPT-6 Sol, has a 1,050,000 token context window and 128,000 maximum output tokens. Those are GPT-6 Sol figures, so check OpenAI's documentation for the GPT-6.1 Sol limits.