Google rolls out Gemini 4 Argon to a small group of cybersecurity partners and says it outperforms GPT-6 Astra on certain coding and knowledge work benchmarks (Madison Mills/Axios)

Google has unveiled Gemini 4 Argon, the company's highly anticipated flagship model, stating that it exceeds expectations across most benchmarks. The new model outperforms OpenAI's GPT-6 Astra and Anthropic's Opus 5.5 and Fable 5.1 by a significant margin, with Argon scoring an impressive 51.3% on AutomationBench, nearly nine points higher than Opus 5.5.

In knowledge work tasks, Argon demonstrates remarkable capabilities, achieving 84.2% on the GraphWalks test for inputs between 256K and 1M tokens, a nearly 12-point advantage over GPT-6 Astra. This model also boasts an exceptional 77.9% score on DeepSWE v1.1, setting a new state of the art in coding abilities, though it lags behind in some areas, such as FrontierSWE v2 and Terminal-Bench 4.0.

Despite its strengths, Argon's performance on coding benchmarks like FrontierSWE v2 and Terminal-Bench 4.0 reveals mixed results, with GPT-6 Astra and Opus 5.5 leading in these categories. One standout feature of Gemini 4 Argon is its ability to generate up to one million output tokens, a significant upgrade from previous Gemini models, which previously limited output to 64,000 tokens.

This expanded capacity allows the model to tackle complex problems in a single reasoning trajectory, providing a deeper level of reasoning. Additionally, Google emphasizes that Argon has been trained to autonomously identify, validate, and patch software vulnerabilities, with the model excelling in internal benchmarks compared to its predecessors.

While Google has not disclosed pricing details for Argon, the announcement highlights the model's potential to significantly enhance productivity for standard office tasks and its promising capabilities in coding and knowledge work.