- Gemini 4 Argon can generate up to one million tokens in a single output.
- Google says the model is already being used internally for coding, research and large-scale engineering work.
- The AI is initially being tested by trusted cyber defenders before wider access.
Google has unveiled Gemini 4 Argon, a new frontier AI model that can generate up to one million tokens in a single output, a sharp increase from the previous 64,000-token limit.
In simple terms, the bigger limit is designed to let the Google AI model work through much longer and more complicated tasks without having to break them into as many separate steps. Google says this could be useful for software development, financial and legal research, enterprise work and other jobs that require sustained reasoning.
The company said Argon is already being used internally by thousands of its employees for debugging, research and large-scale software projects. It has also been used to optimise quantum computing algorithms, improve memory efficiency in Google's data centres and help migrate C and C++ code to Rust.
Google described the model as capable of "deep reasoning across complex, long-horizon workflows", reflecting its focus on tasks that can take much longer than a typical chatbot exchange.
Why the million-token limit matters
The jump from 64,000 to one million tokens is not simply a bigger number. Tokens are the small units into which AI models break text and other information, so a higher output limit gives a model more room to reason through a task and produce a lengthy response.
Google says Argon can use that capacity for large codebases, lengthy research and multi-step enterprise workflows. In one example, the company said Argon replaced 32,000 lines of SIMD code in its libgav1 video decoder and produced a Rust version that was 2.7 times faster than an earlier Rust port while producing the same video output.
The model is also designed to work with visual information, including professional charts, long videos and collections of documents.
On the DeepSWE v1.1 benchmark, which measures performance on long-horizon software engineering tasks, Google reported a score of 77.9 per cent. It also reported a 51.3 per cent score on Zapier's AutomationBench, which assesses end-to-end execution across business functions.
These are Google's own reported benchmark results, rather than independent assessments.
The AI is going after cyber threats too
Cybersecurity is another major focus for Argon. Google says the model can autonomously find, validate and patch critical software vulnerabilities.
The company said Argon uncovered a critical vulnerability affecting healthcare software used by hospitals worldwide, which it said previous frontier models had missed. On CWE-bench v1, a test of vulnerability remediation, Google reported that Argon tied for first place with a score of 68 per cent.
That capability is also one reason Google is not releasing the model to everyone immediately.
Argon is initially being made available to a select group of trusted cyber defenders through Google's Fairwind Program. Google says the phased rollout will give it more time to test safety measures and strengthen safeguards against cyber misuse, prompt injection attacks and other risks.
Google said it will expand access to developers, enterprises and consumers, starting with paid API customers and Google AI Ultra subscribers. It has not given a final date for wider consumer availability.
For businesses, the model will initially cost $2 per million input tokens and $10 per million output tokens, with cached input tokens receiving a 95 per cent discount. Google said the prices will later rise to $4 per million input tokens and $20 per million output tokens.
The bigger question may be what businesses do with an AI model that can sustain much longer workflows. Rather than simply generating a paragraph, writing a short piece of code or answering a single question, Argon is being positioned for jobs where the work itself can stretch across thousands of lines of code, large collections of documents and multiple stages of reasoning.
