Google release ‘Gemini 4 Argon’

Google DeepMind has officially unveiled Gemini 4 Argon, marking the debut of its next-generation Gemini 4 frontier AI model family. Designed as a specialized reasoning engine rather than a conventional conversational assistant, Argon is built to execute long-horizon software engineering tasks, enterprise workflows across legal and financial analysis, and defensive cybersecurity operations.

The headline technical leap is an output token capacity of up to 1 million tokens in a single response—up from the prior 64,000-token ceiling—allowing the model to generate entire software repositories, end-to-end audit reports, and multi-file code refactors in a single execution pass. To manage the safety implications of its defensive vulnerability discovery and autonomous tool execution, Google is implementing a phased rollout, granting initial access to vetted security teams through its Fairwind Program and internal engineering units ahead of general API availability.

Key Takeaways

• First Gemini 4 Generation Model: Gemini 4 Argon represents Google DeepMind’s latest frontier architecture, moving away from standard Pro/Flash naming conventions in favor of the “Argon” designation.
• 1 Million Output Tokens: The model raises maximum single-turn output capacity to 1,000,000 tokens, enabling sustained reasoning loops and single-pass repository generation.
• DeepSWE Benchmark Record: Google reports a 77.9% score on the long-horizon DeepSWE v1.1 software engineering evaluation, outperforming rival frontier checkpoints, though trailing on terminal-based suites like Terminal-Bench 4.0.
• Aggressive Launch Pricing: The model carries introductory pricing of $2.00 per million input tokens and $10.00 per million output tokens (rising to $4.00 and $20.00 post-launch), with a 95% discount on cached inputs ($0.10 per million tokens).
• Restricted Fairwind Access: Initial access is limited to trusted enterprise security partners under the Fairwind Program and Google internal teams, with paid API customers and Google AI Ultra subscribers scheduled for subsequent access phases.
• Internal Google Deployments: Google confirmed Argon agents have already been deployed internally, contributing to data center memory optimizations (freeing over 300 TiB) and automating legacy C/C++ to Rust migrations.

Technical Specifications & Architecture

Gemini 4 Argon is optimized for tasks requiring extended chains of reasoning, cross-document synthesis, and iterative constraint verification:

Specification / Dimension	Gemini 4 Argon	Prior Frontier Standard (Gemini 1.5 / 3.x)	Competitor Reference (Claude Opus / GPT Frontier)
Max Output Capacity	Up to 1,000,000 Tokens	64,000 Tokens	128,000 Tokens
Input Context Window	2,000,000+ Tokens	2,000,000 Tokens	128K – 200K Tokens
Introductory API Pricing	$2.00 In / $10.00 Out (per 1M)	Standard Tier Rates	~$3.00–$5.00 In / $15.00–$25.00 Out
Standard Post-Launch Price	$4.00 In / $20.00 Out (per 1M)	N/A	N/A
Cached Input Discount	95% Off ($0.10 per 1M tokens)	Variable caching rates	50% – 80% Off
Primary Domain Focus	Software Engineering, Cyber, Law	Multimodal general synthesis	General agentic workflows

The 1-Million Output Token Pipeline

While existing frontier models typically cap output generation between 8,000 and 128,000 tokens, Argon’s 1-million-token output window allows developers to run continuous agentic trajectories. Instead of requiring complex orchestration frameworks to split code generation across multiple API turns, Argon can draft complete applications, perform large-scale legacy codebase rewrites, or generate comprehensive technical documentation within a single call.

Benchmark Landscape: Where Argon Leads and Trails

Google published performance evaluations comparing Argon against industry peers across software engineering, business automation, and cybersecurity:

                            BENCHMARK PERFORMANCE SNAPSHOT
                                          │
       ┌──────────────────────────────────┼──────────────────────────────────┐
       ▼                                  ▼                                  ▼
SOFTWARE ENGINEERING (DeepSWE)     ECONOMIC IMPACT (Vals Index)       VULNERABILITY REMEDIATION
• Argon:         **77.9%** (SOTA)  • Argon:         **68.9%** (Rank 1)• Argon:         **68.0%** (Tied 1st)
• Claude Opus:   74.2%             • GPT Frontier:  65.4%             • Rival Systems: 68.0%
• GPT Frontier:  74.1%             • Claude Opus:   64.1%             • Evaluated on CWE-bench v1

• Software Engineering (DeepSWE v1.1): Argon posted 77.9%, establishing a new state of the art on multi-file issue resolution in large code repositories. However, on environments requiring live terminal interactions (Terminal-Bench 4.0), Argon scored 57.4%, trailing Claude Opus’s 66.4%.
• Economic Task Performance (Vals Index v2.1): Across GDP-weighted professional tasks spanning finance, corporate law, tax filing, and coding, Argon ranked first on the leaderboard with 68.9%.
• Defensive Cybersecurity (CWE-bench v1): On vulnerability discovery and remediation, Argon achieved a 68% success rate, tying with custom agent harnesses while demonstrating the capability to identify zero-day vulnerabilities in hospital enterprise infrastructure during red-team trials.

Phased Rollout and Safety Architecture

Because Argon possesses significant vulnerability discovery and autonomous software modification capabilities, Google is following a gated deployment schedule under its Frontier Safety Framework:

                           ARGON DEPLOYMENT ROADMAP
                                      │
     Phase 1: Trusted Cyber Vetting ──► Phase 2: Enterprise API ──► Phase 3: Ultra Subscriptions
      (Fairwind Program & Internal)       (Vertex AI & AI Studio)      (Consumer Pro Interface)

1. Fairwind Program: Early access is restricted to vetted defensive cybersecurity organizations, participating research labs, and Google internal teams. This phase operates with tailored defensive guardrails to allow vulnerability patching without exposing offensive exploitation vectors.
2. Voluntary Pre-Release Review: Google is participating in the US government’s voluntary safety review process to evaluate chemical, biological, radiological, and nuclear (CBRN) as well as autonomous cyber risks.
3. Wider Enterprise Release: General availability via Google Cloud Vertex AI, Google AI Studio, and Google AI Ultra consumer plans will follow as safety reviews and infrastructure capacity scale.

Frequently Asked Questions (FAQs)

What is Google Gemini 4 Argon?

Gemini 4 Argon is Google DeepMind’s first fourth-generation frontier AI model, designed specifically for complex software engineering, enterprise knowledge work (legal/finance), and defensive cybersecurity.

What is the output token limit for Gemini 4 Argon?

Gemini 4 Argon features a maximum single-turn output capacity of up to 1 million tokens, a major increase from previous generation models that were capped at 64,000 tokens.

How much does Gemini 4 Argon cost to use?

Google announced an introductory API rate of $2.00 per million input tokens and $10.00 per million output tokens, with cached input discounted by 95% down to $0.10 per million tokens. Standard pricing will adjust to $4.00 input and $20.00 output after the introductory window.

Who can access Gemini 4 Argon today?

Access is currently limited to trusted cybersecurity defenders participating in Google’s Fairwind Program and internal Google engineering teams. Broad availability across Vertex AI, Google AI Studio, and Google AI Ultra subscribers will roll out in subsequent phases.