Google announced Gemini 4 Argon on 30 September 2026 as its new frontier model. If you run a business, the headline matters less than three questions: what is it good at, when can you use it, and should you change your plans for it?
The short answers: it’s built for long, complicated work such as large code changes, financial and legal research, and security reviews. Most businesses can’t use it yet. And no, you shouldn’t put a project on hold waiting for it.
What is Gemini 4 Argon?
Gemini 4 Argon is Google DeepMind’s new frontier model. It was announced by Koray Kavukcuoglu, SVP at Google DeepMind and Google’s Chief AI Architect, who describes it as built for “complex, long-horizon workflows”: jobs that take many steps and hours of work rather than one quick answer.
Google names three areas where it expects Argon to be used:
- Software engineering: debugging, large codebase migrations and performance work.
- Enterprise knowledge work: research and drafting in finance, law and tax.
- Cyber defence: finding, checking and patching security flaws in software.
What’s actually new
Much longer answers. Argon can write up to 1 million output tokens in one response, up from 64K tokens before. (A token is a small piece of text, often part of a word.) In practice that means a full report, a long migration plan or a large block of code in one go, instead of stitching dozens of replies together.
Strong benchmark results, as Google reports them. Google published these scores for Argon:
| Benchmark | Area | Argon’s score |
|---|---|---|
| DeepSWE v1.1 | Real-world software engineering | 77.9% |
| AutomationBench | Automation | 51.3% (ranked first) |
| LVBench | Understanding long videos | 91.7% |
| CWE-bench v1 | Fixing security weaknesses in code | 68% (tied first) |
Google also says Argon leads the Vals Index, which covers finance, coding, legal and tax work, and Harvey’s Legal Agent Benchmark for legal research and drafting.
Treat these as a signal, not a promise. Every AI company chooses which tests to publish, and these are results Google reported itself. The only score that matters for your business is how a model does on your own tasks.
Real work inside Google. Google says its own teams are already using Argon on large engineering jobs, including moving C and C++ code to Rust (one codebase runs to more than 800,000 lines), speeding up a video decoder, and memory optimisations it expects to free more than 300 TiB across its data centres once rolled out.
Who can use it today
Very few people outside Google. Google is rolling Argon out first to “trusted cyber defenders” through its Fairwind Program. It is not in the Gemini app.
Paid API customers and Google AI Ultra subscribers are next in line, followed by wider access. Google hasn’t given dates. In its own words, “safely releasing frontier capabilities at this level requires a phased approach.”
Why so careful? A model that can find and patch security holes can also help someone exploit them. Google says it has safeguards in four areas: preventing misuse such as cyber attacks, resisting prompt injection, monitoring for misaligned behaviour, and hardened sandbox environments.
Google isn’t alone here. OpenAI says its new GPT-6 Astra meets the “Critical” cyber capability threshold in its own safety framework, and it keeps advanced security work behind a separate access programme.
What it will cost
When Argon opens up through the API, Google’s announced prices are:
| Price | Input, per 1M tokens | Output, per 1M tokens |
|---|---|---|
| Introductory | $2 | $10 |
| After the introductory period | $4 | $20 |
Cached input tokens get a 95% discount. Google hasn’t said how long the introductory period lasts.
For context, the introductory price matches Claude Sonnet 5.5 and GPT-6.1 Sol ($2 in, $10 out), and the later price matches Claude Opus 5.5 ($4 in, $20 out). That’s aggressive pricing for a frontier model.
What this means for your business
- Don’t wait for Argon. If you’re planning an AI feature, such as a support assistant, document processing or a reporting agent, today’s models can handle most of it. Google’s own Gemini 3.8 Flash costs $0.75 per million input tokens and $3.75 per million output tokens at its introductory price (until 31 December 2026), and Google describes it as “engineered for long-horizon software engineering, autonomous agents, and complex enterprise workflows”. Start there and move up only if your tests show you need to.
- Build so you can switch models. Keep your prompts, tools and test cases separate from the model, so changing models is a settings change rather than a rewrite. This isn’t theory. In June 2026, Anthropic had to suspend access to its newest models for nearly three weeks after US export controls were applied to them. Teams that could switch to another model kept working.
- Know where a frontier model earns its price. Large code migrations and refactors, long reports that must stay consistent across hundreds of pages, security reviews of your own code, and analysis of long video. A chatbot answering order-status questions isn’t one of them.
- Build your test set now. Collect 30 to 50 real examples of the task you want to automate, with the answers you’d accept. When Argon, or any new model, opens up, you can test it in a day instead of guessing from benchmark tables.
- Check the data path. Before customer data goes to any model, know which product and terms you’re using, where the data is processed and who can see it.
FAQ
Is Gemini 4 Argon available in the Gemini app?
Not yet. Access is limited to early groups, starting with cyber defenders in Google’s Fairwind Program. Google says Google AI Ultra subscribers and paid API customers come next, without a date.
How much will Gemini 4 Argon cost?
$2 per million input tokens and $10 per million output tokens at the introductory price, then $4 and $20. Cached input is 95% cheaper.
Is Argon better than GPT-6 Astra or Claude?
TechCrunch reports Google’s claim that Argon scored significantly higher than OpenAI’s GPT-6 Astra and Anthropic’s Fable and Opus models across a range of benchmarks. Those are vendor claims on vendor-chosen tests. For your business, the answer comes from testing each model on your own work.
What should we use until then?
For most business software: Gemini 3.8 Flash, Claude Sonnet 5.5 or GPT-6.1 Sol. We compare them in Claude vs Gemini vs GPT in 2026.
Want AI in your product?
We build AI features into business software and test models on a sample of your real data before you commit. See our AI services or book a free 30-minute call.
Sources
- Google: Gemini 4 Argon: our next era of frontier intelligence (30 September 2026)
- Google: The latest AI news we announced in September 2026
- Google: Introducing Gemini 3.8 Flash and 3.8 Flash Cyber
- Google AI for Developers: Gemini models
- TechCrunch: Google releases Gemini 4 Argon, called its most powerful model yet
- OpenAI: GPT-6 Astra
- Anthropic: Redeploying Claude Fable 5