What Gemini 4 Argon Means for Google’s AI Race

Google announced Gemini 4 Argon on September 30 as the flagship model for its Gemini 4 generation. It is built for complex, extended work in software engineering, legal and financial tasks, and cybersecurity—but, for now, Google is rolling it out to selected cyber defenders rather than the public. 9

The announcement is also a response to a difficult stretch for Google’s flagship-model plans. Argon is larger than the company’s previous advanced Pro models, and Google’s own benchmark comparisons show strong results against OpenAI and Anthropic systems. But its restricted release means most developers and businesses cannot yet test those claims for themselves. 6

What Argon is designed to do

Google describes Argon as its most capable model so far for complex workloads. Its stated focus includes long-running software-engineering tasks, enterprise knowledge work such as legal and finance work, and cybersecurity defense. The model is larger than Google’s earlier top-tier Pro models, but the company has not disclosed a parameter count in the reporting available here. 9

That makes “larger” a relative description, not a precise measure of Argon’s size. It also does not, on its own, establish how the model compares with rivals in cost, speed, or performance on a user’s specific task.

What Google’s benchmark comparisons show

Google reports strong results on several tests relevant to coding and professional work. In the company’s published comparison, Argon scored 68.9% on the Vals Index for knowledge work, versus 63.1% for OpenAI’s GPT-6 Astra and 67.0% for Anthropic’s Claude Opus 5.5. On AutomationBench, Google reports 51.3% for Argon, compared with 41.4% for Astra and 42.5% for Opus 5.5.

Google also reports a 77.9% score on DeepSWE v1.1, a test of long-horizon software engineering tasks. Reporting on the cybersecurity comparisons describes Argon as tying for first on a cybersecurity measure, rather than establishing an across-the-board lead. 37

These figures are useful evidence of what Google says Argon can do, but they should be read as benchmark results—not as a settled verdict on overall model quality. The tests cover selected tasks, and broad access is still restricted, limiting independent evaluation by most users. 6

Who can use Argon now?

Google is initially providing Argon to selected, vetted cybersecurity defenders through its Fairwind Program. The company says the limited rollout is intended to help defenders use the model while reducing the risk of misuse. Google has also joined the U.S. government’s voluntary process for pre-release model access. 65

Argon is not yet broadly available to developers, businesses, or consumers, and the reporting cited here gives no firm date for general access. 6 That cautious launch resembles Anthropic’s restricted access to its Claude Mythos Preview, according to reporting on the rollout; the available sources do not provide enough detail for a full comparison of the two programs. 6

How the Gemini 3.5 Pro delay fits in

Google’s earlier Gemini 3.5 Pro plans were delayed while the company worked to improve the model’s capabilities, particularly coding, according to reporting citing people familiar with the matter. Google subsequently moved to Gemini 4 Argon rather than releasing 3.5 Pro.

That history raises the stakes for Argon: Google is presenting it as a new flagship after a period in which rivals OpenAI and Anthropic advanced their own frontier models. But the available reporting attributes the 3.5 Pro delay to capability work; it does not substantiate a specific DeepMind organizational change as the cause.

A frontier-model push, with cost still an open question

Google has also emphasized the cost advantage of its models as part of its competitive positioning. That gives the company another possible way to compete alongside headline benchmark performance. But the sources cited here do not establish a verified cost advantage for Argon itself over rival flagship models.

The clearest current picture is therefore mixed: Google is claiming strong results on selected coding and enterprise benchmarks, while limiting access as the model is evaluated. Whether those results translate into an advantage in everyday use, and how Argon’s cost compares with competitors, remain open questions until broader testing is possible. 6