Gemini 4 Argon: 1M Output Tokens and Deep Reasoning

Quick answer

• Gemini 4 Argon is Google DeepMind's frontier reasoning model engineered for long-horizon software engineering, enterprise domain workflows (legal, finance, tax), and defensive cybersecurity.
• The model expands the maximum generation headroom to an industry-record 1 million output tokens (up from 64K), allowing autonomous agents to deliberate, plan, and generate vast codebases in a single continuous trajectory.
• In software development, Argon achieves a state-of-the-art 77.9% on DeepSWE v1.1. Within Google, it is currently driving autonomous C/C++ to safe Rust refactoring across projects like libgav1 and the Fuchsia Zircon kernel.
• In defensive cybersecurity, Google has partnered with organizations like Wiz (via the Scan for Good initiative) and is providing unguardrailed access to verified cyber defenders to discover and patch critical infrastructure vulnerabilities.
• Argon launches with an introductory API price of $2 per million input tokens and $10 per million output tokens, with cached input tokens discounted by 95% ($0.10 per million).

Image source: Google Official Blog (The Keyword) / Google DeepMind

As artificial intelligence shifts from casual conversational chatbots to autonomous coding agents and enterprise workflows, the demands placed on foundation models have fundamentally changed. Simple benchmark retrieval and short-response synthesis are no longer the bottleneck. Today's challenge is long-horizon reasoning: maintaining coherence across complex code refactoring, auditing massive legal documents, and simulating multi-step security attack surfaces.

With Gemini 4 Argon, Google DeepMind introduces its most specialized frontier reasoning architecture to date. Here is a technical breakdown of how Argon works, its real-world engineering benchmarks, and its architectural significance.

The 1-million output token breakthrough

Most commercial frontier models cap output generation between 4,096 and 65,536 tokens. While input context windows have scaled into the millions, constrained output budgets forced developers to stitch together multiple agent loops, fragile recursive summarization steps, and chunked code generation passes.

Gemini 4 Argon expands the output token limit to 1,000,000 tokens:

• Single-Trajectory Deliberation: Rather than prematurely truncating deep reasoning steps, Argon has the headroom to explore multiple computational hypotheses, test internal edge cases, and self-correct before emitting final artifacts.
• Whole-Project Generation: Developers and software architects can task the model with rewriting entire modules, generating comprehensive unit and integration test suites, or producing multi-hundred-page technical documentation in a single prompt execution.
• Extended Video and Multimodal Ingestion: Paired with its 2M+ input context window, Argon leads the LVBench benchmark with a 91.7% score for long-form video comprehension and document-level reasoning.

Real-world engineering: Autonomous C/C++ to Rust migration

Rather than evaluating models solely on competitive programming riddles, Google tested Argon internally across critical production infrastructure.

The model is actively participating in fleet-wide modernization programs:

1. Large-Scale Kernel and Library Refactoring: Argon agents are translating mission-critical C/C++ libraries (such as the re2 regex engine and libgav1 video decoder) to memory-safe Rust, scaling up to the 800,000+ line Fuchsia Zircon kernel.
2. Compiler-Guided Vectorization in libgav1: When porting Google's open-source AV1 video decoder (libgav1), Argon replaced 32,000 lines of legacy SIMD assembly. By running profile-guided experiments and inspecting compiler assembly output, it generated safe Rust that auto-vectorizes cleanly, delivering a 2.7x speedup over previous human-written Rust ports while producing byte-identical video frames.
3. Datacenter Memory Fleet Optimization: Autonomous Argon agents analyzed Google telemetry to identify systemic memory leaks and allocation inefficiencies, freeing over 300 TiB of RAM with projected savings up to 1 PiB.

On DeepSWE v1.1, an evaluation framework measuring real-world multi-file software engineering tasks, Gemini 4 Argon established a new state of the art with a 77.9% resolve rate.

The table below highlights Argon's performance across key frontier benchmarks.

Defensive cybersecurity: The Fairwind Program and Wiz collaboration

Cybersecurity represents one of the most dual-use domains in artificial intelligence. A model capable of writing zero-day exploits can also defend infrastructure, provided it is deployed into the hands of defenders.

Google is rolling out Argon to vetted security organizations through its Fairwind Program:

• Unguardrailed Defensive Access: To allow defenders to perform realistic penetration testing and code audits, verified security researchers receive access without standard conversational refusals that typically hinder vulnerability analysis.
• Real-World Discovery with Wiz: Through Wiz's Scan for Good initiative, Argon uncovered critical zero-day vulnerabilities in hospital and healthcare management software deployed globally, preventing exposure of patient records before malicious actors could weaponize the flaws.
• Automated Patch Verification: On CWE-bench v1, Argon tied for the top score at 68%, demonstrating that it not only identifies Common Weakness Enumerations (CWEs) but writes mathematically sound regression patches that preserve codebase semantics.

Robustness against indirect prompt injection

A fundamental vulnerability of agentic systems connected to external tools, web searches, and user inboxes is Indirect Prompt Injection (IPI), where adversarial text hidden in emails or web pages hijacks model intent.

Argon introduces hardened architectural safeguards:

• Gray Swan IPI Leadership: The model leads benchmark rankings on Gray Swan's Indirect Prompt Injection evaluation through extensive automated red teaming and adversarial reinforcement learning.
• Chain-of-Thought Monitoring: Google deploys internal monitors that scrutinize the hidden chain-of-thought activations during agent execution, pausing runtime if the model drifts from the operator's original constraints.
• Hermetic Agent Sandboxing: Evaluated agents run in isolated, sealed microVM environments adhering to Google DeepMind's Agent Control Roadmap.

API pricing and deployment roadmap

Google has announced an aggressive pricing structure designed to encourage high-volume agentic adoption:

• Input Tokens: $2.00 per million tokens.
• Output Tokens: $10.00 per million tokens.
• Context Caching: $0.10 per million tokens (a 95% discount), dramatically reducing costs for long-running agents that repeatedly reference large codebases or repository trees.

Initial access is rolling out via private preview to Fairwind security partners and enterprise testers, with general availability planned across Google AI Studio, Vertex AI, and Google AI Ultra subscribers.

Sources and further reading

• Google DeepMind, "Gemini 4 Argon: Our Next Era of Frontier Intelligence", Google The Keyword Blog (September 30, 2026).
• Google DeepMind Research Team, "Gemini 4 Argon Model Card and Frontier Safety Evaluations".
• Wiz Research, "AI-Powered Defensive Auditing: Findings from the Scan for Good Initiative with Gemini Argon".
• Vals AI, "The Vals Index: Measuring GDP-Weighted Economic Output of Frontier Models".