Google Unveils Gemini 4 Argon with 1 Million Output Tokens and Strong Performance

Google has finally unveiled Gemini 4 Argon, its latest large language model (LLM) that boasts impressive performance on complex workloads. The pretraining of Gemini 4 began on July 21, after the performance of Gemini 3.5 fell short of expectations due to leadership changes at Google DeepMind and a major restructuring of the research organization.

The new model will initially be made available to select cybersecurity partners before being rolled out to paid API customers and Google AI Ultra subscribers in phases. Google is taking a cautious approach, emphasizing its commitment to safety through a self-regulatory process for early access to models by the US government.

Gemini 4 Argon has achieved strong performance on various benchmarks, including DeepSWE v1.1, where it set a new high of 77.9%, and ranked third on Artificial Analysis' Coding Agent Index. The model's maximum output token limit has been expanded from 64,000 to 1 million tokens, making it one of the industry's highest levels.

Google also highlighted Argon's potential for cybersecurity defense capabilities, with the company planning to provide a version without security guardrails for verified security teams to use. The model's cost per task is $1.99, which is relatively inexpensive compared to its performance.