Google has officially revealed its latest flagship artificial intelligence model named Gemini 4 Argon. The new architecture expands the output ceiling to 1000000 tokens while outscoring rival models in key software engineering benchmarks. Access is rolling out to Google AI Ultra subscribers and paid API customers, though advanced security features remain restricted to vetted specialists.
The largest architectural change centers on token generation. Gemini 4 Argon can generate up to 1000000 output tokens in a single response, replacing the earlier ceiling of 64000 tokens. This expansion allows the system to inspect sprawling software projects, locate code defects, and execute chained multi step instructions without restarting the prompt cycle.
In the DeepSWE v1.1 engineering benchmark, Gemini 4 Argon reached an accuracy score of 77.9%. That result places it ahead of competing frontier models like GPT 6 Astra at 74.1% and Claude Opus 5.5 at 74.2%. The system also demonstrated high accuracy across legal analysis, financial research, and long video processing tasks.
Google trained the model on vulnerability detection and code security. In the CWE bench v1 evaluation, the model scored 68%, matching existing flagship tiers across the industry. Google already runs Argon based agents internally to optimize server infrastructure and translate legacy software code from C and C++ into Rust.
General access is rolling out to paid developers using the official API alongside Google AI Ultra subscribers. Full access to the specialized cyber defense tools remains closed to the general public. Google is restricting those specific capabilities to authorized security partners and internal engineering teams.
