SILICON VALLEY — Google has released its most advanced artificial intelligence model after a seven-month gap, giving it an opening to catch up with OpenAI and Anthropic. Inside the company, however, employees are raising questions about how well the model actually performs.
Google unveiled its next-generation AI model, Gemini 4 Argon, on Sept. 30. The model succeeds Gemini 3, released last November, and is the company's first top-tier release in about seven months, following Gemini 3.1 Pro in February. At its developer conference in May, Google said it would launch Gemini 3.5 Pro in June, but it abandoned that release and moved straight to the next version.
Google said Argon delivers strong performance in knowledge-intensive professional work such as finance, law and taxation, as well as in areas including cybersecurity. To let the model tackle complex problems, Google expanded its output token limit to 1 million tokens, the highest in the industry, from 64,000 tokens.
The new model outperformed OpenAI and Anthropic models on coding and task-performance benchmarks, and matched OpenAI's most advanced model, GPT-6 Astra, in cybersecurity assessments. In an intelligence index compiled by AI model evaluator Artificial Analysis, Argon tied for third place with Anthropic's Claude Fable 5.1 and GPT-6 Astra.
Google stressed that it had trained the model to defend against cyberattacks, in an apparent nod to recent controversy over hacking of AI agents. Argon can autonomously identify and patch critical software vulnerabilities.
Unease persists within Google, however. Bloomberg, citing a person inside the company, reported that Argon's benchmark scores came out well but that employees are running into difficulties when using it for coding work. Experts said Google may be struggling with "benchmaxing," a focus on benchmark scores at the expense of real-world performance.
