Google Gemini 4 Argon: What Its Benchmarks, 1M-Token Limit and Limited Rollout Mean

Google announced Gemini 4 Argon on September 30 as the flagship model of its new Gemini 4 generation. It is designed for long-running software engineering, professional knowledge work and cybersecurity defense. The launch also marks the arrival of Google’s next frontier model after months of delays—but Argon is not yet a broadly available public release. 11114

What Google says Argon can do

Google positions Argon for complex workflows in areas including software engineering, legal and financial work, and cyber defense. The emphasis is on handling tasks that take multiple steps or require sustained work, rather than only answering short, standalone prompts. 11

Google says its engineers are already using the model for work such as debugging and codebase migrations. Those internal uses offer an early example of the kinds of tasks the company expects Argon to handle, though they are not independent evaluations of its performance. 310

Benchmark results: strong claims, with a caveat

Google reports a score of 77.9% on DeepSWE v1.1, a benchmark for long-horizon software engineering. It also says Argon achieved 68% on CWE-bench v1, tying for first place on a benchmark related to remediating security vulnerabilities. 91022

These figures are Google’s reported results. They indicate performance on particular benchmarks, not a universal ranking across models, coding tasks or real-world security work. The announcement and reporting available here do not establish that Argon is best at every task. 1022

What the 1-million-token output limit changes

Argon’s maximum output rises from 64,000 tokens to 1 million tokens, according to Google’s announcement and coverage of the launch. That gives the model room to generate much longer responses in a single output—potentially useful for tasks that produce extensive code or detailed work products. 3422

The distinction matters: this figure describes how much the model can produce, not how much text or code it can accept as input. A larger output limit makes longer generations possible; by itself, it does not show that the model reasons more accurately or performs better on every task.

Who can use Gemini 4 Argon now?

At launch, access outside Google is limited to vetted cybersecurity defenders through the Fairwind Program. Google has described the restricted rollout as a phased approach for a model with advanced cyber-defense capabilities. 21011

Google plans to widen access to paid API customers and Google AI Ultra subscribers. The reporting available does not give a firm date for that broader rollout, so Argon should be understood as announced and entering limited access—not generally available to the public. 5623

Why the launch matters

Argon is significant both as Google’s next frontier model and as a controlled release after delays to its next-generation AI plans. Google is emphasizing coding and cybersecurity benchmarks to position it against rival models, while limiting initial access to a group of cyber defenders. 114

For now, the clearest takeaways are the model’s intended use, its unusually high output ceiling and Google’s benchmark claims. How those results translate to broader use will be easier to assess once access expands and independent users can evaluate the model across more tasks.