Gemini 4 Argon and 1 million tokens in one response: pricing and use cases
Gemini 4 Argon writes 1 million tokens in a single response. Who has it, what it costs, where it loses and how to prepare.
Gemini 4 Argon 1 million tokens is the headline number of Google DeepMind's new frontier model, introduced on September 30, 2026: it can write up to one million tokens in a single response, up from 64 thousand before. Almost nobody can use it yet: only trusted cybersecurity defenders get it. Here are the pricing, the results and what to prepare while it opens up.
For the short summary of the launch and the restricted access, read the Gemini 4 Argon summary and its defenders-only access.
Who can use Gemini 4 Argon today
Google describes it as its frontier model for coding, enterprise knowledge work and cybersecurity. It did not open it to the public: trusted defenders receive it inside a program called Fairwind. According to reports, that is more than 650 organizations, including governments, critical infrastructure operators and security companies.
The reason, per Google, is that those defenders get it with the cybersecurity filters turned off so it can find vulnerabilities and fix them without refusing. That is why the company says releasing capabilities at this level requires phases, with internal and external red teams and monitoring. Wiz, one of the security companies, already used it and found critical flaws in healthcare software that earlier models missed. The next announced step: paid API customers and Google Ultra subscribers, in that order, with no date.
What 1 million tokens in a single response means
Maximum output went from 64 thousand tokens to one million: sixteen times more. A token is roughly three quarters of a word, so a million is about 750 thousand words in one response. Before, you had to request work in pieces and make sure they fit together; now the question is how big the assignment can be.
Approximate conversions (they come from that rule, not from Google):
- A 500-word page is about 670 tokens: a million is about 1,500 pages.
- A 100-word email is about 130 tokens: about 7,500 emails.
- A 300-page book is about 200 thousand tokens: five books fit.
- An average line of code is about 10 tokens: about 100 thousand lines, a large application.
Use them to size a project, not to invoice it.
How much Gemini 4 Argon costs
The launch price is 2 dollars per million input tokens and 10 per million output tokens. Afterward it rises to 4 and 20. The announcement reviewed gives no end date for the introductory price.
- A maximum one-million-token response: 10 dollars today and 20 later.
- Reading 500 thousand tokens of code: 1 dollar of input today and 2 later.
- With caching, which discounts input by 95%, re-reading it costs 5 cents today and 10 later.
- A full pass: 11 dollars now and 22 when the introductory period ends.
Results: where it wins and where it loses
Per Google, Argon leads most of the benchmarks it published. On DeepSWE v1.1 it scores 77.9%, versus 74.2 for Opus 5.5 and 74.1 for Astra. On the Vals index, 68.9 versus 67 and 63.1. On long context, GraphWalks between 256 thousand and one million tokens: 84.2 versus Astra's 71.8. Real advantages, but not overwhelming.
What doesn't make the headlines: on FrontierSWE v2 Argon scores 55.0 and Astra 65.5, so it loses by 10.5 points. It also trails on scientific terminal tasks. And on Harvey's legal exam it scores 19.6%: first place, but the best model still fails most of the test.
Two benchmarks matter more if you sell services. AutomationBench (business process automation): 51.3%, first place, meaning it still fails nearly half of those processes. CVE-Bench (fixing known vulnerabilities): 68%, tied for first, about two out of three. The useful reading: it works as a helper that moves work forward, not yet as an employee you leave a process to.
How to apply it in your business?
These are our own analysis proposals, not Google examples, and all of them end in a draft someone must review.
- Migrating a store on an old system. Gather code and documentation (about 500 thousand tokens), ask for the plan and new files in one response, run the tests and send back only the errors; caching makes the second read cheaper. The first pass costs about 11 dollars of model. If a developer charges 50 dollars an hour, a day is 400 and 11 is under 3%. That's an assumption: use your own rate.
- Financial or legal research. Gather the annual reports of five companies, ask for a comparative memo with page citations and check each citation against the original. The announcement doesn't say how much input text Argon accepts, so this is a hypothesis, not a promise.
- Long video. On LVBench, the long-video test, it scores 91.7% (a Google figure). A consultancy could ask for a summary of eight-hour meetings with agreements, owners and the exact minute. Video also spends tokens: measure a real meeting before promising a rate.
If your work is writing emails, summarizing a short PDF or drafting posts, a small model does it for cents and today. Argon pays off when the assignment is huge. Prepare at no cost: pick a real job and measure how many tokens it is, note what it costs you today in hours and money, and calculate with both prices (2 and 10, and 4 and 20).
Risks and limits
Defenders get it unfiltered to find flaws and fix them; Google insists it is for defensive use only. That is the tension of every security tool: port scanners were born to defend and attackers use them too. The public version will likely arrive with filters and feel more cautious than the benchmarks suggest. And every number is published by Google; there is no independent verification yet.
Frequently asked questions
What is Gemini 4 Argon?
It is the frontier model Google DeepMind introduced on September 30, 2026, built for coding, enterprise knowledge work and cybersecurity.
Can I use Gemini 4 Argon now?
No, for now only trusted defenders get it through the Fairwind program. Paid API customers and Google Ultra subscribers come next, with no date.
How much does it cost?
At launch, 2 dollars per million input tokens and 10 per million output tokens. Afterward, 4 and 20.
How much is a million tokens?
About 750 thousand words, around 1,500 pages or about 100 thousand lines of code.
Conclusion
Gemini 4 Argon and its 1 million token output open the door to huge assignments in a single response, but today it is closed, its launch price won't last and it loses exactly where few people mention. The advantage will go to whoever already has what to feed it when it opens. Tell me in the video comments: what project of yours would fit in a million tokens?