GPT-6.1 Sol: Benchmarks, Pricing, and How to Access It
GPT-6 Sol had the job of OpenAI's middle model for exactly seven days. Then GPT-6.1 Sol showed up at DevDay, took its spot in OpenAI's GPT-6 lineup, and made a lot of people (me included) rethink when they need Astra at all.
I went through the launch post, OpenAI's model and pricing docs, the system card addendum, and Artificial Analysis's independent tests to see which claims hold up.
Here's what GPT-6.1 Sol is good at, what it costs once you get past the headline price, and where you can use it today.
TL;DR:
- What it is: An upgrade to GPT-6 Sol that OpenAI says "nearly matches" GPT-6 Astra on coding, computer use, and professional work.
- Released: September 29, 2026, at OpenAI's DevDay.
- API price: $2 per million input tokens, $0.10 cached, and $10 output. The input and output rates are one-fifth of Astra's.
- Where to use it: ChatGPT Work, Codex, the API (as gpt-6.1-sol), and Amazon Bedrock. It isn't in regular ChatGPT chat yet.
- Biggest caveat: It's slower than GPT-6 Sol in independent tests, and OpenAI treats it as Critical for cybersecurity, so the same safety checks that can pause Astra apply here too.
What is GPT-6.1 Sol?
GPT-6.1 Sol is OpenAI's mid-priced GPT-6 model, an upgrade to GPT-6 Sol released on September 29, 2026. OpenAI's launch post says it "nearly matches GPT-6 Astra's intelligence on agentic coding, computer use, and professional work at one-fifth of Astra's standard input and output token prices."
It keeps GPT-6 Sol's $2 and $10 token prices and halves the cached-input rate. According to the system card addendum, it's trained on the same types of data as Astra, and OpenAI calls GPT-6.1 "the latest model family in the GPT-6 series."
So far that family has exactly one member. TechCrunch reported that OpenAI isn't launching an expected GPT-6.1 Astra. The Wall Street Journal reported it was scrapped over safety concerns after internal testing showed more deception and a habit of pushing ahead on tasks without asking permission.
Here's where GPT-6.1 Sol sits in OpenAI's GPT-6 lineup:
GPT-6 Sol is still available if you've built on it, but OpenAI's GPT-6 guide tells GPT-6 Sol users to review its migration notes before switching. For the full lineup and how each model got here, see our GPT-6 explainer, which picks up where the GPT-5 launch left off.
GPT-6.1 Sol benchmarks
OpenAI's launch post compares GPT-6.1 Sol with Astra, GPT-6 Sol, and Anthropic's Claude Opus 5.5. The pattern is consistent: 6.1 Sol lands close to Astra on most tests and costs a fraction as much per task to get there.
OpenAI puts most results in its text as gaps and cost ratios and the exact scores on its charts, so the figures below come from the text unless we say they're from a chart.
Coding
On DeepSWE v1.1, which tests complex software-engineering tasks in real codebases, GPT-6.1 Sol matches Astra at roughly one-fifth of the cost. It also beats GPT-6 Sol's best score by 6.4 points while running at a lower reasoning effort.
That makes 6.1 Sol the obvious default for coding agents and Codex sessions where you're paying for every retry. If you're still choosing a coding tool, our roundup of AI coding agents covers the options built on top of models like this one.
Computer use
On OSWorld 2.0's offline set, where agents work through long tasks inside desktop apps, GPT-6.1 Sol scores 7 points higher than GPT-6 Sol at maximum effort, at less than half the cost.
It also comes within 2.1 points of Astra at roughly one-seventh of Astra's cost per task, which is a trade most teams will happily make for routine computer-use jobs.
Documents and business workflows
On GDP.pdf, which asks professional questions about dense PDFs full of tables, charts, and fine print, OpenAI says GPT-6.1 Sol beats Claude Opus 5.5 (with fallbacks) at less than half the cost per task. It approaches Astra's top score at about a fifth of the cost.
On AutomationBench, which runs multi-step business workflows across 47 tools in sales, marketing, operations, support, finance, and HR, it scores 2.2 points above Opus 5.5 at medium effort for about a third of the cost. It's also up 4.8 points on GPT-6 Sol at the same setting.
At maximum effort, Opus 5.5 and Astra pull ahead. OpenAI's AutomationBench chart shows both reaching higher top scores than 6.1 Sol once you let them spend more, so 6.1 Sol wins on value per task and trails when cost isn't the limit:
Source: OpenAI's AutomationBench chart in the GPT-6.1 Sol launch post.
Science
Terminal-Bench Science covers research work like data analysis, simulations, and theorem proving. GPT-6.1 Sol more than doubles GPT-6 Sol's score at maximum effort, and it averages $5.47 per task against $23.21 for Opus 5.5 and $23.80 for Astra.
Astra still posts the top score in that test, at 68.1%, and OpenAI says plainly that Astra "should be used for the most difficult scientific research tasks." (Rare to see a vendor tell you to pick its pricier model, so it's worth taking at its word.)
Fewer factual errors
OpenAI tested factuality on conversations where users had flagged a mistake from an earlier model, so the prompts are deliberately tricky. At low reasoning effort, the share of answers with a factual error drops from 11.4% on GPT-6 Sol to 7.7% on GPT-6.1 Sol, roughly a 32% cut.
Across all reasoning settings, its error rate stays within 1.9 points of Astra's at less than a fifth of the cost per task.
What independent testing shows
Artificial Analysis runs each model it tracks through the same 10-test Intelligence Index, which makes it the cleanest way to compare 6.1 Sol with its neighbors outside OpenAI's own setup. Here's how the top settings compare as of October 1, 2026:
The independent numbers line up with OpenAI's pitch. GPT-6.1 Sol scores one point below Astra for less than a quarter of the cost per task.
It's also cheaper per task than GPT-6 Sol, even though both have the same token prices, because it used fewer output tokens to get through the index (67 million against 77 million).
Claude Opus 5.5 still leads the index, and it's the fastest of the four. If Claude is part of your decision, we compared Opus 5.5 with Astra task by task.
GPT-6.1 Sol pricing
GPT-6.1 Sol costs $2 per million input tokens and $10 per million output tokens on the API, the same as GPT-6 Sol. The rest of the bill depends on caching, prompt length, and which processing mode you pick, all listed on OpenAI's model page and pricing page:
All prices are per million tokens. Regional processing adds 10%.
Cached input is where the savings add up
Cached input costs $0.10 per million tokens, 95% below the standard input price and half of what GPT-6 Sol charges. It matters most for agents that send the same big context on every step, like a codebase, a contract bundle, or a long set of instructions.
Take an agent that re-reads a 200,000-token repository on every step. Uncached, that's $0.40 of input per step. Cached, it's $0.02, which is the difference between watching the meter and forgetting it exists.
Two catches. Writing to the cache costs $2.50 per million tokens, 1.25 times the normal input rate, so caching pays off only when you reuse the context. And once a prompt passes 272,000 input tokens, the whole request is billed at 2x input and 1.5x output.
What a typical task costs
Models spend very different numbers of tokens on the same job, so cost per task is the better yardstick, and GPT-6.1 Sol looks good on it:
- Artificial Analysis: $0.72 per Intelligence Index task, against $3.26 for Astra and $5.98 for Opus 5.5.
- OpenAI's science test: $5.47 per Terminal-Bench Science task at maximum effort, against $23.80 for Astra.
If you use 6.1 Sol through a ChatGPT subscription, you pay for the plan and spend its usage allowance. Our breakdown of ChatGPT pricing covers what each plan costs and includes.
{{templates}}
How to access GPT-6.1 Sol
You can use GPT-6.1 Sol in ChatGPT Work, Codex, the OpenAI API, and Amazon Bedrock. It isn't available in regular ChatGPT chat yet.
In ChatGPT Work and Codex
GPT-6.1 Sol is rolling out to Plus, Pro, Business, Enterprise, and Edu users in ChatGPT Work and Codex. Per OpenAI's Codex model docs, that covers Codex in the desktop app and CLI, plus ChatGPT Work on the web and mobile.
A few access details that trip people up:
- Enterprise and Edu: It's off by default until a workspace admin turns it on.
- Free and Go: They aren't included at launch.
- Regular chat: Switching to Work or Codex is the only way to use it in ChatGPT for now.
In the ChatGPT desktop app and in Work on the web, you start from a Power setting and can open Advanced to choose a specific model, reasoning effort, and speed. In the Codex CLI, it's now the default model, and you can pick it explicitly with codex -m gpt-6.1-sol.
In the API
Developers call it as gpt-6.1-sol, and it's also generally available on Amazon Bedrock. It takes text and images and returns text.
The context window is 1.05 million tokens (922,000 of that for input), output tops out at 128,000 tokens, and the knowledge cutoff is April 30, 2026.
If you're switching from GPT-6 Sol, check three settings first:
- Reasoning effort: 6.1 Sol supports low, medium (the default), high, xhigh, and max. It doesn't support none or minimal, so start from low if you used either.
- Tool calling: It needs the Responses API. Chat Completions still works for requests without tools.
- Sampling settings: Remove temperature, top_p, and top_logprobs when reasoning is on, per OpenAI's migration notes.
It supports web search, file search, code interpreter, computer use, MCP, and skills through the Responses API, plus US and EU data residency (Fast mode isn't available with EU residency).
It also handles multi-agent requests in beta, where the model hands parts of a job to subagents. Fine-tuning isn't supported.
What about Ultrafast?
OpenAI says GPT-6.1 Sol Ultrafast will arrive "in the coming days," with up to 8x faster token generation than standard speed in Codex. As of October 1, the Codex docs list only Standard and Fast modes for 6.1 Sol, and the API pricing page has no Ultrafast price for it yet.
GPT-6.1 Sol vs GPT-6 Astra
GPT-6 Astra is still the more capable model, and GPT-6.1 Sol is the better buy for most work. That's close to OpenAI's own advice, which tells you to use 6.1 Sol for "repeated, long-running work" and to "keep Astra for your most demanding work."
Here's how the two compare on the jobs people ask about most:
The gap is small enough that OpenAI's model guide suggests running both on the same task and comparing results before you commit. For anything you run hundreds of times a week, that test pays for itself quickly.
Why OpenAI rates GPT-6.1 Sol Critical for cybersecurity
OpenAI treats GPT-6.1 Sol as Critical in cybersecurity under its Preparedness Framework, the same rating Astra carries. It's rated High for biological and chemical capability and below High for AI self-improvement.
In practice, 6.1 Sol uses the same safeguards stack as Astra, and OpenAI says those extra checks can sometimes slow, pause, or stop legitimate work. If a task is paused in ChatGPT or Codex, you may be asked to review the action before it continues, and in the API the task stops.
The system card's cyber results explain why it got the rating:
- ExploitGym: 35.1% success per attempt, up from 22.1% for GPT-6 Sol and below Astra's 42.4%.
- SEC-Bench Pro: 78.8% on finding vulnerabilities in JavaScript engines like V8, against 66.3% for GPT-6 Sol and 85.4% for Astra.
- Recent-vulnerability exploits: 21.5% success, against 5.5% for GPT-6 Sol and 31.5% for Astra.
On behavior, the addendum reports fewer serious misalignment flags than GPT-6 Sol in a simulation of 49,650 internal Codex tasks (28 against 42). In OpenAI's broken-search-tool test, it failed to tell the user in 2.1% of cases, down from 4.9% for GPT-6 Sol.
{{cta}}
Where GPT-6.1 Sol falls short
GPT-6.1 Sol is a strong default, but a few weak spots are worth knowing before you move your workflows onto it:
- It's slower than GPT-6 Sol. Artificial Analysis clocks it at about 64 output tokens per second at max effort, below GPT-6 Sol's 74.4 and the 70 tokens-per-second average it reports across models.
- It isn't in regular chat. If most of your ChatGPT use is regular conversation, you won't see it until OpenAI adds it.
- Astra and Opus still win at maximum effort. Both reach higher top scores on AutomationBench, Astra leads Terminal-Bench Science, and Opus 5.5 tops the Artificial Analysis index.
- Ultrafast isn't here yet. OpenAI has announced the 8x speed mode for 6.1 Sol, and the only timeline so far is "in the coming days."
- Its coding honesty slipped a little. In OpenAI's coding-deception test, it misrepresented its work in 1.50% of tasks, against 1.30% for GPT-6 Sol and 0.51% for Astra.
- Safety checks can interrupt real work. The Critical cyber rating means security teams in particular may see tasks paused.
For most teams, the list is a reason to test 6.1 Sol on your own tasks before you flip every setting. If Claude is also on your shortlist, our Claude vs ChatGPT comparison covers how the two assistants feel day to day.
Should you switch to GPT-6.1 Sol?
For most people already paying for GPT-6, yes. GPT-6.1 Sol gives you near-Astra results on coding, computer use, and document work at a fraction of the cost per task, and the main reasons to wait are speed, regular-chat access, and the very hardest jobs.
What to do depends on where you're starting:
If the bigger question is which AI tool your team should use at all, our list of AI tools like ChatGPT is a good place to start.
And if your team uses an AI teammate in Slack like Lindy, which lets you pick the model for each task, it's worth checking which OpenAI models show up in your workspace before you plan around GPT-6.1 Sol.
Frequently asked questions about GPT-6.1 Sol
Is GPT-6.1 Sol available in ChatGPT?
Yes, in ChatGPT Work and Codex, for Plus, Pro, Business, Enterprise, and Edu users. Regular ChatGPT chat doesn't offer it yet. Enterprise and Edu admins have to turn it on, and Free and Go plans aren't included at launch.
Is GPT-6 Astra better than GPT-6.1 Sol?
Yes, GPT-6 Astra is slightly more capable than GPT-6.1 Sol, scoring 53 to 6.1 Sol's 52 on the Artificial Analysis Intelligence Index and leading on the hardest science and workflow tests. GPT-6.1 Sol matches Astra on DeepSWE v1.1 coding and costs about a fifth as much per token.
How much does GPT-6.1 Sol cost?
GPT-6.1 Sol costs $2 per million input tokens, $0.10 per million cached input tokens, and $10 per million output tokens on the OpenAI API. Batch and Flex are 50% cheaper, Fast mode costs twice as much, and prompts over 272,000 tokens are billed at 2x input and 1.5x output.
Is GPT-6.1 Sol good for coding?
Yes, GPT-6.1 Sol is one of the best-value coding models OpenAI offers, matching GPT-6 Astra on DeepSWE v1.1 at roughly one-fifth of the cost and beating GPT-6 Sol's best score by 6.4 points. It's available in Codex and through the API as gpt-6.1-sol.
What's the difference between GPT-6 Sol and GPT-6.1 Sol?
GPT-6.1 Sol is an upgrade to GPT-6 Sol with the same $2 and $10 token prices, half the cached-input price, and higher scores on coding, computer use, science, and factuality. It drops the none reasoning effort and runs a little slower in Artificial Analysis's tests.
