Google has made Gemini 4 Argon available to enterprise users and cloud developers, positioning the model for long, multi-step tasks spanning software development, document analysis, finance, legal work and cybersecurity.
The model is available in Gemini for business as well as through Google’s developer platforms under the identifier gemini-4-argon. Google says Argon is designed to maintain context while working across extended tasks and mixed inputs such as text, tables, charts and images.
Enterprise workflows are the focus
Google is emphasizing completed work products rather than short conversational answers. Examples in the company’s announcement include reconciling ledgers, analyzing variances, synthesizing long documents and updating several files during a software project.
The company says the model was trained to respect system boundaries, report negative findings and pause before irreversible actions. Those are important characteristics for agent-style systems, but organizations will still need their own permissions, approval gates, audit trails and evaluations before allowing a model to take consequential steps.
Google cited early-access feedback from partners in development, legal services, workflow automation and design. These accounts are useful indicators of intended use, but they remain vendor-selected results rather than independent benchmarks of every production environment.
Cloud access and pricing
Enterprise developers can use Argon through Google’s Gemini platform and Google AI Studio. Google also lists availability through its Antigravity development environment.
Introductory API pricing is set at $2 per million input tokens and $10 per million output tokens through November 30, 2026, with cached input priced at a discount. Google says standard pricing will then rise to $4 per million input tokens and $20 per million output tokens.
The temporary rate gives teams a limited window for evaluation, but cost tests should model the later price as well. Long-running agents can repeatedly send large contexts, use tools and generate intermediate outputs, causing real workflow costs to differ sharply from a simple prompt-and-response calculation.
Controlled access for cybersecurity testing
Google is also offering a version without cyber guardrails to selected government, critical-infrastructure and cybersecurity participants through its limited-access Fairwind Program. That access is separate from the standard commercial model and is intended for vetted defensive testing.
For most organizations, the immediate task is validation: measuring accuracy, task completion, citation quality, latency and cost on real internal workflows. A capable model can reduce manual effort, but reliable deployment depends on the surrounding controls as much as the model itself.
Source: Google Cloud: Gemini 4 Argon available on Google Cloud.
Featured image: AI-generated editorial illustration by Bex News.
