- Gemini 4 Argon is Google’s new frontier AI model, announced on 30 September 2026. Google calls it the start of its “next era of frontier intelligence”.
- You can’t use it yet. For now it is only open to selected cybersecurity defenders through Google DeepMind’s Fairwind Program and to Google’s own staff. Wider access will start with paid API customers and Google AI Ultra subscribers; Google has not given a date.
- Headline numbers: it can write up to 1 million tokens in one answer (up from 64,000), and Google reports a record 77.9% on DeepSWE v1.1, a software-engineering test.
- Launch price for developers: $2 per million input tokens and $10 per million output tokens, rising to $4 / $20 after an introductory period.
Google has announced Gemini 4 Argon, the first model of the Gemini 4 generation and the most capable model the company has described so far. It arrives in one of the busiest weeks of the year for AI launches: Anthropic released Claude Sonnet 5.5 on 28 September and OpenAI released GPT-6.1 Sol at its DevDay on 29 September. This guide explains what Argon is, what Google says it can do, who can use it and what it will cost, in plain language.
What is Gemini 4 Argon?
Gemini 4 Argon is a large reasoning model from Google DeepMind. In its announcement, Google describes it as built “to sustain deep reasoning across complex, long-horizon workflows” — in other words, long jobs with many steps, rather than quick one-question chats. Google names three areas where it is strongest:
- Software engineering — writing, debugging and migrating large amounts of code.
- Enterprise knowledge work — professional research and drafting in fields such as law and finance.
- Cybersecurity defence — finding, checking and fixing security holes in software.
Google also mentions strong multimodal skills (understanding charts, documents and long videos) and creative writing. The name follows Google’s newer naming style: “Argon” is the model’s name within the Gemini 4 family, much as OpenAI now uses Astra, Sol and Luna for GPT-6.
Can I use Gemini 4 Argon now?
Not yet, unless you work for one of the organisations Google has invited. Google says it is taking a “phased approach” because of how capable the model is:
- Now: a group of trusted cyber defenders gets access through the Fairwind Program, and thousands of Google employees already use Argon internally. Google says it is also taking part in the U.S. government’s voluntary process for pre-release model access.
- Next: Google says it will release Argon “as soon as possible” to developers, enterprises and consumers, starting with paid API customers and Google AI Ultra subscribers.
- Later: Google has not said when, or whether, Argon will reach the free Gemini app or the cheaper Google AI plans.
If you are wondering whether it’s worth paying for Ultra to get early access, our guide Is Gemini free? Google AI Plus, Pro and Ultra plans explained covers what each plan includes today. Since there is no public release date yet, upgrading only for Argon would mean paying for something you may not get for a while.
What is the Fairwind Program?
Fairwind is Google DeepMind’s early-access programme for cybersecurity defenders. According to its programme page, it is aimed at governments and national cyber authorities, critical-infrastructure operators (such as healthcare, telecoms, energy and finance), core technology platforms and academic labs. Members must use strong login protection, including phishing-resistant multi-factor authentication, and may only give Argon to their internal security, incident-response or penetration-testing teams. It is not a waitlist for ordinary users.
What Google says Gemini 4 Argon can do
Much longer answers: 1 million output tokens
The most unusual change is the output limit. Most chatbots can take in a lot of text but only write a limited amount back in one go. Google says Argon can produce up to 1 million tokens in a single answer, up from 64,000 for earlier models. (A token is a chunk of text roughly three-quarters of a word long, so a million tokens is several long novels’ worth.) Google’s argument is that when the model has room to “think deeply and generate hundreds of thousands of tokens” in one run, it can solve hard problems in one go instead of stopping halfway.
For everyday users this matters less than it sounds: nobody wants a million-word reply to a question. It matters for long automated jobs, such as rewriting a large codebase or producing a full research report with all its working.
Coding
Google reports that Argon sets a new state of the art on DeepSWE v1.1, a test of long, real-world software-engineering tasks, with 77.9%. It also gives examples of how Google uses it internally:
- Argon agents are moving C and C++ code to the memory-safe Rust language across Google, from libraries of tens of thousands of lines up to more than 800,000 lines for the Fuchsia Zircon kernel. Google says these rewrites are still being audited and reviewed before they reach production.
- For libgav1, Google’s open-source video decoder, Argon rewrote 32,000 lines of low-level code in an existing Rust version; Google says the result runs 2.7 times faster than that Rust version with identical video output.
- A team of Argon agents analysed memory use across Google’s data centres and applied optimisations that Google says freed more than 300 TiB of memory.
Office and professional work
Google says Argon is the leading model on the Vals Index, which weighs results in finance, coding, law and tax work by each sector’s share of the U.S. economy, and that it ranks first on Zapier’s AutomationBench, a test of multi-step business workflows, with 51.3%. It also reports leading results on Vals Finance Agent v2 and Harvey’s Legal Agent Benchmark.
Understanding video and documents
On LVBench, which tests understanding of long videos, Google reports a state-of-the-art 91.7%. Google says Argon can analyse professional charts, pick out details from long videos and act on a series of documents.
Cybersecurity
This is where Google puts the most emphasis. It says Argon can “autonomously find, validate, and patch critical software vulnerabilities”. The numbers Google has published:
| Test (as reported by Google) | Gemini 4 Argon | Gemini 3.8 Flash Cyber |
|---|---|---|
| CWE-bench v1 (fixing vulnerabilities) | 68% (tied for first) | — |
| Real-world vulnerability discovery | 85.8% | 71.0% |
| Wiz penetration-test benchmark | 70.9% | 58.2% |
Google also says the security company Wiz used Argon in its free Scan for Good programme and found a critical vulnerability in healthcare software used by hospitals, one that “previous frontier models had missed”. For trusted defenders and its own teams, Google says it is releasing Argon without its usual cyber guardrails.
As always, these are figures the company chose to publish about its own model. Independent testing will show how Argon compares once more people can use it.
How much will Gemini 4 Argon cost?
For developers using the API, Google has announced an introductory price of $2 per million input tokens and $10 per million output tokens, with cached input 95% cheaper. Once the introductory period ends, the price becomes $4 input and $20 output. Google has not said how long the introductory period lasts.
At the introductory rate, Argon costs the same as two other models released the same week — OpenAI’s GPT-6.1 Sol and Anthropic’s Claude Sonnet 5.5 — and at the full rate it matches Claude Opus 5.5. All three are compared side by side in Gemini 4 Argon vs GPT-6.1 Sol vs Claude Sonnet 5.5. To see what a real workload would cost, try the AI API cost calculator or browse prices in the AI models database.
For consumers, pricing will depend on which Google AI plan includes Argon. So far Google has only said the Ultra plan will be among the first.
Safety: why Google is holding it back
Google spends a large part of the announcement on safety, which explains the slow rollout. It says it is strengthening safeguards in four areas before broad release:
- Misuse: Argon is designed to refuse help with cyber, chemical, biological, radiological and nuclear attacks while still allowing legitimate research. Google says it now also monitors the model’s internal activity to spot misuse.
- Prompt injection: attacks where hidden instructions in a web page or document try to hijack the AI. Google calls Argon its most resistant model yet and says it leads Gray Swan’s indirect prompt injection benchmark.
- Misalignment: systems that watch Argon’s reasoning and actions and stop it if it tries to go beyond what the user asked.
- Hardened test environments: sealed sandboxes for risky training and evaluation.
This caution matches the wider mood this week: OpenAI said it would not release GPT-6.1 Astra after internal safety tests raised concerns (our report), and released the smaller GPT-6.1 Sol instead.
What Gemini 4 Argon means for everyday Gemini users
- If you use the free Gemini app: nothing changes today. Google has not announced Argon for free users.
- If you pay for Google AI Ultra: you are first in line for consumer access, but there is no date yet. Watch the model menu in the Gemini app.
- If you build apps on Gemini: paid API customers are also early in the queue. The launch price is attractive for a frontier model, but budget for the higher price after the introductory period.
- If you just want the best AI today: Argon is not an option yet. Claude Sonnet 5.5, GPT-6.1 Sol and Gemini 3.8 Flash are all available now. New to Gemini? Start with our beginner’s guide to Gemini.
Frequently asked questions
When was Gemini 4 Argon released?
Google announced it on 30 September 2026. On that date it went only to trusted cyber defenders through the Fairwind Program; the public release has no date yet.
Is Gemini 4 Argon free?
No free access has been announced. Google says the rollout will start with paid API customers and Google AI Ultra subscribers.
How much does the Gemini 4 Argon API cost?
The introductory price is $2 per million input tokens and $10 per million output tokens, with cached input at 95% off. After the introductory period it rises to $4 and $20.
What is special about Gemini 4 Argon?
According to Google: an output limit of 1 million tokens (up from 64,000), a state-of-the-art 77.9% on the DeepSWE v1.1 coding test, first place on AutomationBench (51.3%) and the Vals Index, and strong cybersecurity-defence skills.
Is Gemini 4 Argon better than GPT-6.1 Sol or Claude Sonnet 5.5?
On the tests Google chose, Argon posts leading scores, but each company reports different benchmarks under different settings, so the numbers can’t be lined up directly. Sonnet 5.5 and GPT-6.1 Sol have one big advantage: you can use them today.
What is Gemini 3.8 Flash Cyber?
It is the earlier Google model built for cybersecurity work, which Google uses as the comparison point for Argon’s security results.