Michael JohnsonOn September 30, Google introduced Gemini 4 Argon—a model for software engineering, enterprise tasks, and cyber defense. At first, it is being made available to a limited group of cybersecurity specialists. The company promises to expand access, but independent assessment of its claimed capabilities remains difficult for now.
In its announcement, Google describes Argon as a model for complex, long-running workflows. Google said it plans to make it available later to developers, businesses, and consumers, starting with paid API customers and Google AI Ultra subscribers. The company did not specify a timeline for expanding access.
Google announced introductory pricing of $2 per million input tokens and $10 per million output tokens; cached input tokens will cost 95% less. After the introductory period, according to a footnote in the announcement, prices will be $4 and $20, respectively. These are announced rates for a future launch, not confirmation that the model is already available to everyone.
The company also specifies a limit of up to 1 million output tokens. This is a generation limit, not the size of the context window.
What the Internal Results Show
Google gives examples of Argon’s use in its own projects. According to the company, the model helped optimize a quantum algorithm, exceeding a published baseline result by 40%. Google also reports that implementing optimizations found by agents in its data centers freed more than 300 TiB of memory; the company estimates potential overall savings at 500 TiB–1 PiB.
In another example, Google says a Rust version of the libgav1 video decoder runs 2.7 times faster than the previous Rust port and produces identical output. The company describes these cases as internal engineering work. On their own, they do not prove that Argon is generally superior to other models or will deliver the same results for external teams.
The announcement also includes results from coding and cybersecurity tests. These are figures published by Google itself; without independent replication and comparable testing conditions, they should be treated as company claims rather than independent confirmation of the model’s superiority.
Why the Access Sequence Matters
Google says the phased rollout is necessary to continue testing and strengthen safeguards before wider distribution. This approach is particularly important for a model the company positions as a cybersecurity tool. But limited access alone proves neither that Argon is unusually dangerous nor that its claimed capabilities are real.
So far, the announcement and Google’s stated access plan are confirmed. How reproducible the internal results are outside the company can be assessed after access expands and independent evaluations become available.