GPT-6.1 Sol is OpenAI's upgrade to GPT-6 Sol, released on September 29, 2026, and OpenAI says it nearly matches GPT-6 Astra on agentic coding, computer use and professional work at one-fifth of Astra's standard token prices. It costs $2 per million input tokens and $10 per million output tokens, and you can use it today in ChatGPT Work, Codex and the API as gpt-6.1-sol.
📺 Watch: GPT-6.1 Sol Just Made a HUGE Breakthrough in Coding!
🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside
That's the whole story in two sentences. Astra-level work, Sol-level price.
I covered the launch on my channel the day after it dropped. I run Hermes every day with a separate profile per model, so a new middle-tier model that punches this close to the top tier changes what I put where. Here's what changed, where you can get it, the numbers that matter and who should switch.
If you want the exact profiles, prompts and model routing I use, they're inside the AI Profit Boardroom for $69/mo, with daily tutorials every time a model like this drops.
What's new in GPT-6.1 Sol vs GPT-6 Sol
GPT-6 Sol only came out on September 22. One week later, 6.1 replaced it. That's fast, even for OpenAI. In my launch video I called it out: Sol got swapped out after seven days.
So what did a week of extra work buy you? Four things.
1. A big jump in quality at the same price
Standard input and output prices didn't move. You still pay $2 in and $10 out per million tokens. But the scores went up across almost every benchmark OpenAI published. On Terminal-Bench Science, OpenAI says 6.1 Sol more than doubles GPT-6 Sol's score at maximum effort, at less than half the cost per task.
2. Cached input is half the price
Cached input now costs $0.10 per million tokens. OpenAI says that's 95% less than standard input and 50% less than GPT-6 Sol's cached price. If you run agents that reuse the same big context over and over, this is where your bill shrinks.
3. Fewer factual errors
OpenAI tested it on deliberately hard prompts, taken from conversations where users had flagged an earlier model's mistake. At low reasoning effort, the share of answers with a factual error dropped from 11.4% to 7.7%. That's roughly a 32% cut. Across settings, OpenAI says it stays within 1.9 points of Astra's error rate at less than a fifth of the cost per task.
4. Better behaviour as an agent
OpenAI says 6.1 Sol is more transparent about its limits and better at respecting user intent and safety constraints. In its tests it failed less often than GPT-6 Sol at admitting a broken search tool, respecting explicit restrictions and avoiding unauthorised outcomes in agentic tasks.
One thing got taken away. Per OpenAI's developer guide, GPT-6 Sol supported a "none" reasoning effort. GPT-6.1 Sol doesn't. The lowest setting is low. If your app uses none, switch it to low when you migrate.
Where GPT-6.1 Sol is available
Here's where you can use it right now, according to OpenAI.
- ChatGPT Work: Plus, Pro, Business, Enterprise and Edu users. Web and mobile.
- Codex: the same plans, in the desktop app and the CLI.
- API: model ID gpt-6.1-sol, through the Responses API, Chat Completions (without tool calling) and Batch.
- Not in Chat yet: OpenAI says it isn't available in regular ChatGPT Chat.
Two catches. Free and Go plans aren't included at launch. And on Enterprise and Edu, an admin has to switch it on because it's off by default.
Then there's Ultrafast. OpenAI promised up to 8x faster token generation than standard speed in Codex. The API model page lists Ultrafast at 6x standard pricing, which works out to $12 in and $60 out per million tokens. In Codex, OpenAI's docs say Ultrafast is on the Pro $500 plan and eligible Enterprise and Edu plans.
Not sure which plan or tier makes sense for your business? Grab a free AI strategy session and we'll map out the models and workflows that actually pay for themselves.
📺 Watch: Qwen 3.8-Max, Fable 5 and GPT-5.6 Sol: Who Wins?
GPT-6.1 Sol specs and benchmarks
The spec sheet from OpenAI's model page is simple. A 1,050,000-token context window. Up to 922,000 input tokens and 128,000 output tokens. Text and image in, text out. Knowledge cutoff April 30, 2026. No audio or video.
It plugs into the full tool set on the Responses API: web search, file search, image generation, code interpreter, hosted shell, apply patch, skills, computer use, MCP and tool search.
Reasoning effort goes low, medium, high, xhigh and max. Medium is the default.
Now the benchmarks. These are OpenAI's own published numbers, best score per model, with cost per task where OpenAI charted it.
| Benchmark | GPT-6.1 Sol | GPT-6 Sol | GPT-6 Astra |
|---|---|---|---|
| DeepSWE v1.1 (coding) | 75.2% at high, $0.65/task | 68.8% at max, $2.74/task | 74.1% at xhigh, $4.43/task |
| OSWorld 2.0 offline (computer use, max) | 71.4%, $1.27/task | 64.4%, $3.37/task | 73.5%, $9.44/task |
| AutomationBench (medium) | 31.7% | 26.9% | 34.1% |
| Terminal-Bench Science (max) | 57.0%, $5.47/task | 27.6% | 68.1%, $23.80/task |
| Factual error rate (low, lower is better) | 7.7% | 11.4% | within 1.9 pts |
Look at that DeepSWE row. 6.1 Sol's best score is higher than Astra's best score, and the task costs a fraction of the price. That's the headline.
OpenAI also compared it against Claude. On AutomationBench at medium effort, 6.1 Sol beat Opus 5.5 by 2.2 points at roughly a third of the cost. On GDP.pdf, which tests answering professional questions from complex PDFs, OpenAI says it beat Opus 5.5 with fallbacks at less than half the cost per task.
Where Astra still wins: science. OpenAI is clear that Astra keeps the top score on Terminal-Bench Science at 68.1% and should be used for the hardest research tasks.
Benchmarks are a starting point. I run my own head-to-heads on real business tasks in Goldie Bench, because a model that wins a chart can still trip over your actual workflow.
📺 Watch: GPT-6 + Doug: OpenAI's Secret AI
What I saw when I tested it
In my launch video I put it through one-prompt 3D world builds. The worlds were playable and passable, but less polished than what I got from Sonnet 5.5. So for flashy front-end visuals, it's not my first pick.
Where it shines is the work you run all day. Agent loops, coding tasks, document jobs, multi-step automations. Astra can burn through your token allowance fast. Sol at a fifth of the price means you can keep going.
The coding side is the bigger story, and OpenAI has since published more on it. I broke that down separately in my GPT-6.1 Sol coding guide, including the COBOL migration result.
Who GPT-6.1 Sol is for
OpenAI's own guide puts it neatly. Astra is the highest intelligence. Luna is the fastest and cheapest. Sol is the balance. If you want the full breakdown of the other two, read GPT-6 Sol vs GPT-6 Luna.
Here's how I'd put it.
- Pick 6.1 Sol if you run complex work often. Coding agents, computer-use agents, research over long documents, business automations. OpenAI's advice is the same: Astra when maximum quality matters most, Sol for complex work you run all the time.
- Pick Astra for the hardest one-off problems, especially scientific research.
- Pick Luna for high-volume, simple tasks where speed and cost beat everything.
If you're running Hermes, I showed how I wire the GPT-6 models in my Hermes GPT-6 walkthrough, and the whole setup sits in my Agent OS guide.
Want the plug-and-play version? The AI Profit Boardroom gives you the Agent OS, a 30-day AI agent roadmap, four coaching calls a week and a prompt library, with 3,400+ members building alongside you.
How to start using it today
In ChatGPT, open Work mode and pick GPT-6.1 Sol from the model picker. In Codex, run codex --model gpt-6.1-sol or set model = "gpt-6.1-sol" in your config.toml.
In the API, set the model to gpt-6.1-sol on the Responses API. Use Responses if you need tool calling. Start at medium effort, then move up only when the task needs more planning.
If you'd rather have someone build the plan with you, book a free AI strategy session and we'll look at where this model fits in your business.
FAQ
Is GPT-6.1 Sol better than GPT-6 Astra?
Not across the board. OpenAI says it nearly matches Astra on coding, computer use and professional work, and on DeepSWE its best score edges past Astra's. Astra still leads on scientific research and on OSWorld 2.0 by 2.1 points.
How much does GPT-6.1 Sol cost?
In the API, $2 per million input tokens, $0.10 per million cached input tokens and $10 per million output tokens. Batch and Flex are 50% cheaper. Prompts over 272K input tokens are billed at higher rates for the whole request.
Can I use GPT-6.1 Sol on the free ChatGPT plan?
No. OpenAI says Free and Go aren't included at launch. It's on Plus, Pro, Business, Enterprise and Edu, in ChatGPT Work and Codex, not in Chat.
What is the GPT-6.1 Sol context window?
1,050,000 tokens, with up to 922,000 input and 128,000 output tokens, per OpenAI's model page.
Want help turning models like this into real time saved and real revenue? Book your free AI strategy session today.
