Google DeepMind announced Gemini 4 Argon on September 30, calling it its most powerful model so far. The company says it was built for sustained, deep reasoning across complex, long-running workflows, and its published results put it ahead of rivals on most of the benchmarks it chose. There is a catch: at launch, almost nobody outside Google can use it.
Who can use it
Argon is rolling out in stages. The first group is a small set of trusted cyber defenders in Google’s Fairwind program, who in some cases get the model without the usual cybersecurity guardrails so they can test attack detection and system hardening. Google says paid Gemini API customers and Google AI Ultra subscribers come next, but it has not given a date. Until then, anyone outside the program is limited to reading about the model.
What it costs
Google lists introductory pricing of $2 per million input tokens and $10 per million output tokens, with cached input discounted by 95 percent. The company says the standard price will be $4 and $20, but it has not said when the introductory period ends. At the standard rate Argon costs about the same as Claude Opus 5.5.
The model’s specifications
Argon has a context window of one million tokens. Google also reports a very large output limit, up to one million tokens, which is meant for long jobs such as big code changes or long reports. Other trackers list a lower maximum output at the moment, so the limit is worth confirming when API access opens.
Google’s benchmark results
Google published results on 18 benchmarks, comparing Argon with OpenAI’s GPT-6 Astra and with Anthropic’s Claude Fable 5.1 and Opus 5.5. It leads on most of them. Its reported scores include 77.9 percent on DeepSWE v1.1, which tests long software engineering tasks, 51.3 percent on AutomationBench, which tests business workflows, 65.4 percent on the Vals Finance Agent benchmark, and 91.7 percent on LVBench, a long-video understanding test.
It does not win everywhere. On agentic coding tests such as FrontierSWE v2 and Terminal-Bench 4.0, it trails GPT-6 Astra and Claude Opus 5.5, with 55.0 and 58.4 percent. As always with company-chosen benchmark tables, the selection matters as much as the scores.
Independent early results
Outside evaluators have started to weigh in, mostly on the model’s launch preview. Artificial Analysis gives Argon a score of 53 on its Intelligence Index, level with GPT-6 Astra and one point ahead of GPT-6.1 Sol. At discounted launch prices, it says the model costs about $1.99 per task on that index, roughly 60 percent of what GPT-6 Astra costs, and it notes that Argon is Google’s first proprietary model above the Flash class in over seven months. Vals AI ranks the model first on its own index, with 68.9 percent, ahead of the Claude models it tracks.
Those results come from a limited number of testers, and the discount will end, so the cost picture will change.
Why the cyber angle matters
Giving the first access to security teams reflects a trend among AI labs. Models that are very good at reasoning and coding are also good at finding and exploiting software flaws, so labs have started releasing them to defenders first. Argon scored well on a cybersecurity benchmark in independent tracking, which fits the approach, and Google says it will add safeguards before wider release.
What to watch
The questions to follow are simple. When will paid API access open, and with what rate limits and regions? Will independent tests on real workloads match the launch benchmarks? And what will the price be when the introductory discount ends? Until then, Argon is a promising result without much to try.
What this means
Google has a credible new top-tier model and a price that undercuts its rivals at the start. Whether that turns into a lasting advantage depends on access, because a model that only a few teams can use cannot compete for developers’ attention.
