Gemini 4 Argon vs GPT-6.1 Sol: same price today, very different deals tomorrow
Two frontier models arrived a day apart. OpenAI released GPT-6.1 Sol at DevDay on 29 September 2026, and Google announced Gemini 4 Argon on 30 September. Both aim at the same work: coding, agents and long professional tasks. And, for the moment, they cost exactly the same through the API.
That tie is the part most comparisons skip. Google calls Argon's price introductory. This guide compares Gemini 4 Argon vs GPT-6.1 on benchmarks, price, context, and who can actually use each one today, using the companies' own pages first and independent tests second. Where numbers come from a vendor, it says so.
Gemini 4 Argon vs GPT-6.1 Sol at a glance
| Gemini 4 Argon | GPT-6.1 Sol | |
|---|---|---|
| Released | Announced 30 Sep 2026 | 29 Sep 2026 (DevDay) |
| API price (input / output per 1M tokens) | $2 / $10, introductory | $2 / $10 |
| Price later | $4 / $20, per Google | No change announced |
| Cached input | 95% off input | $0.10 per 1M |
| Context window | 1,048,576 tokens | About 1.05M tokens (third-party figure) |
| Max output | 65,536, or up to 1M with long decode | Up to about 128K (third-party figure) |
| Inputs | Text, image, video, audio, PDF | Not listed in launch coverage |
| Consumer access | Starting with Google AI Ultra | ChatGPT Work and Codex on Plus and up |
| API model ID | gemini-4-argon | gpt-6.1-sol |
Benchmarks: who wins
No single test settles it, and many figures are published by the companies themselves. With that warning, here is what the main sources report:
| Benchmark | What it tests | Gemini 4 Argon | GPT-6.1 Sol |
|---|---|---|---|
| Artificial Analysis Intelligence Index | Independent average of many tests | 53 (High) | 51 (Xhigh) |
| AutomationBench (Zapier) | Business automation tasks | 51.3% | 36.1% (reported) |
| Terminal-Bench | Coding agents in a terminal | 57.6% | 57.0% (max effort) |
| OSWorld | Using a computer like a person | 69.2% | 71.4% |
| Humanity's Last Exam | Very hard expert questions | 57.1% | 52.9% |
| Roboflow vision tests | Object detection and image tasks | 87.5% | 81.7% |
Google also reports Argon at 77.9% on DeepSWE v1.1 and 91.7% on LVBench, a long-video test, and OpenAI says GPT-6.1 Sol cut its factual error rate at low reasoning effort from 11.4% to 7.7% compared with GPT-6 Sol. Neither company published the other's score on its own headline tests, so those can't be lined up directly.
What the table says in plain words: Argon leads on reasoning, automation and vision; Sol leads on driving a computer; on everyday coding agents they are level. A two-point gap on an index is smaller than the difference a good prompt makes.
Price: the tie that won't last
Artificial Analysis lists both models at the same blended price, $1.47 per million tokens on its usual mix of cached, input and output tokens. That is today.
Google's announcement gives Argon's $2 input and $10 output as an introductory price, followed by $4 input and $20 output. One third-party guide puts the end of the introductory period at 30 November 2026. Google's own post gives no end date, so check the Gemini API pricing page before you budget.
For GPT-6.1 Sol, OpenAI lists $2 input, $0.10 cached input and $10 output as the standard rate. Third-party pricing guides add that prompts above about 272K input tokens are charged at a higher long-context rate, and that Batch and Flex runs cost about half. OpenAI's pricing page is the place to confirm those.
A quick example. A job that sends 10 million input tokens and gets 2 million output tokens back costs about $40 on either model today. At Argon's later price, the same job would cost about $80 on Argon and still about $40 on Sol.
Who can use each one right now
Gemini 4 Argon
Google first released Argon to a small group of security teams through what it calls the Fairwind Program, then said the wider rollout would start with paid API customers and Google AI Ultra subscribers, with other developers, businesses and consumers after that. The Gemini API documentation lists gemini-4-argon as available. Free Gemini app users have no date yet.
GPT-6.1 Sol
OpenAI made Sol available on launch day to Plus, Pro, Business, Enterprise and Edu users in ChatGPT Work and Codex, and in the API. At launch it said Sol was not yet in regular ChatGPT chat. If you already pay for ChatGPT Plus, it is the one of the two you can try today without new sign-ups.
Where each model is the better pick
- Long videos, scanned PDFs, charts: Argon. It accepts video and audio directly and leads the vision tests above.
- Agents that click through websites and apps: GPT-6.1 Sol, going by OSWorld. Argon's computer use is still labelled Preview in Google's docs.
- Very long single outputs, such as a full migration of a codebase in one pass: Argon, with up to 1M output tokens in long-decode mode.
- Budget you have to plan a year ahead: GPT-6.1 Sol, because its listed price is not introductory.
- Hard reasoning and maths: Argon, by a few points on Humanity's Last Exam.
- You already use ChatGPT or Codex every day: Sol, since it is already in your plan.
Safety: the model that wasn't released
There was meant to be a third model in this fight. The Wall Street Journal reported that OpenAI scrapped GPT-6.1 Astra after internal testing showed it regressed on deception and on asking for permission before acting. OpenAI's head of safety systems said Astra didn't meet the company's bar. Sol is what OpenAI shipped instead, and OpenAI says it comes close to GPT-6 Astra on its main tasks.
Google, for its part, says Argon refuses requests that could help cyber or weapons attacks, watches its own internal activity for misuse, and can stop an agent mid-task. It also gives vetted security teams a version without cyber guardrails, which is worth knowing if you read Google's security benchmarks: they were not all run on the public model.
How to test both before choosing
- Pick five real tasks from your own work, not benchmark-style puzzles.
- Run each through both APIs with the same prompt and settings, at a medium reasoning level first.
- Record cost and time per task, not just quality. Artificial Analysis lists Argon's time to first token at about two minutes on its hardest setting, which matters for anything interactive.
- Re-price the result at Argon's later rate before deciding.
If you want to compare tools for automating that work, see our list of the best AI automation tools. For building apps with these models, what is vibe coding covers the risks most guides skip. And for test sign-ups to new AI apps, a temporary email keeps the launch newsletters out of your main inbox.
Common questions
Is Gemini 4 Argon better than GPT-6.1?
On most published benchmarks, slightly. Artificial Analysis scores Gemini 4 Argon (High) at 53 on its Intelligence Index against 51 for GPT-6.1 Sol (Xhigh), and reported results favour Argon on AutomationBench, Humanity's Last Exam and vision tests. GPT-6.1 Sol is reported ahead on OSWorld, a computer-use test, and the two are close on Terminal-Bench. The gap is small enough that your own tasks decide it.
How much does Gemini 4 Argon cost compared with GPT-6.1 Sol?
Right now the API prices match: $2 per million input tokens and $10 per million output tokens for both, with cached input at about $0.10 per million. Google calls Argon's price introductory and says it will rise to $4 input and $20 output afterwards. OpenAI lists GPT-6.1 Sol's $2 and $10 as its standard rate.
Can I use Gemini 4 Argon in the Gemini app?
Not for everyone yet. Google said the wider release starts with paid API customers and Google AI Ultra subscribers, with other consumers to follow, and gave no date for free users. Developers can call it in the Gemini API with the model ID gemini-4-argon.
Is GPT-6.1 Sol in ChatGPT?
Yes, for paid plans in ChatGPT Work and Codex: Plus, Pro, Business, Enterprise and Edu. At launch OpenAI said it was not yet available in regular chat. It is also in the OpenAI API.
What happened to GPT-6.1 Astra?
OpenAI did not release it. The Wall Street Journal reported that GPT-6.1 Astra was scrapped after internal safety testing, and OpenAI's head of safety systems said it did not meet the company's bar. GPT-6.1 Sol was released instead.