Google DeepMind Announces Gemini 4 Argon with One Million Output Token Limit

Architecture and Generation Capacity

Google DeepMind announced Gemini 4 Argon on September 30, 2026, as its latest frontier reasoning model. The primary architectural update is an expansion of the generation limit, increasing from 64,000 output tokens in previous models to 1,000,000 tokens in a single execution pass.

This output volume is engineered for uninterrupted multi-step reasoning. The model targets three specific domains: production software engineering, corporate legal and financial analysis, and defensive cybersecurity operations.

API pricing published by Google includes:

• Input tokens: 2.00 dollars per one million tokens.

• Output tokens: 10.00 dollars per one million tokens.

• Input caching: 0.10 dollars per one million cached tokens (representing a 95 percent discount).

Published Benchmark Scores

According to evaluation tables released by Google, Gemini 4 Argon ranked first in 13 out of 19 standardized industry tests:

• DeepSWE v1.1: Scored 77.9 percent on resolving production software engineering issues.

• CWE-bench v1: Reached 68.0 percent for automated vulnerability remediation.

• AutomationBench: Placed first with 51.3 percent on workflow automation tasks.

• LVBench: Scored 91.7 percent on long-duration video understanding.

Restricted Access via the Fairwind Program

Google restricted the initial deployment of Gemini 4 Argon instead of making it immediately available to the public. Access is currently limited to vetted defensive organizations through the Fairwind Program, which includes government entities, healthcare systems, and critical telecommunication providers.

Authorized partners deploy Argon alongside Google's CodeMender agent to analyze security flaws, simulate attacks, and generate automated software patches. Within the Fairwind environment, standard cyber guardrails are removed to allow deep analysis of zero-day exploits.

Wider availability for commercial API customers and Google AI Ultra subscribers is scheduled to follow in subsequent deployment phases.