Google Gemini 4 Argon Released: 1M Output Tokens, DeepSWE SOTA — How It Compares for Coding and Cyber Defense

TL;DR

Google DeepMind launched Gemini 4 Argon on Sept 30 — the first model of the Gemini 4 generation. Three breakthroughs: 1M max output tokens per response (rivals cap at 128K), a new DeepSWE v1.1 SOTA at 77.9%, and introductory pricing of $2/$10 — half of Claude Opus 5.5. But access is restricted: currently available only to cybersecurity partners through the Fairwind Program.

Why 1M Output Tokens Matters

Current frontier model output limits compared:

This means developers can complete large-scale refactors or generate long reports in a single response without splitting across turns. The cost: a full 1M output tokens costs $10 at introductory pricing, $20 after.

Google has not disclosed Argon's input context window size.

Benchmarks: Where Argon Leads and Trails

Google compared Argon against GPT-6 Astra, Claude Opus 5.5, and Claude Fable 5.1 across 18 benchmarks. Argon leads outright on 12 and ties for first on 1.

Where Argon leads:

Where Argon trails:

Artificial Analysis reports that using discounted prices, Argon matches GPT-6 Astra on the Intelligence Index at 60% of the cost per task.

Pricing Comparison

At introductory pricing, Argon's output cost is 1/5 of GPT-6 Astra's.

Cyber Defense: Find, Validate, Patch

Argon is trained to autonomously find, validate, and patch critical software vulnerabilities. Trusted defenders and internal Google teams receive it without cyber guardrails.

On CWE-bench v1 (vulnerability remediation), Argon ties for first at 68%.

Real-world case: Wiz is already using Argon through its Scan for Good initiative. The model found a critical vulnerability in healthcare software used by hospitals worldwide — one that previous frontier models had missed.

Google is strengthening safeguards in 4 areas before broad release:

1. Misuse defenses for cyber and CBRN risks (including activation monitoring under its Frontier Safety Framework)
2. Indirect prompt injection resistance (Argon leads Gray Swan's IPI benchmark)
3. Misalignment monitoring of chain-of-thought and actions (with ability to stop execution)
4. Sealed, isolated sandboxes for high-risk training and evaluations

Google's Internal Results

Thousands of Googlers already use Argon. Google shared 4 internal results:

1. Memory optimization: Agents freed over 300 TiB across data centers, with 500 TiB to 1 PiB projected
2. SIMD code replacement: Replaced 32K lines of SIMD code in libgav1 Rust port; decoder runs 2.7x faster with identical output
3. C/C++ to Rust migration: Up to 800K+ lines in the Fuchsia Zircon kernel
4. Quantum algorithm: Beat a published quantum algorithm baseline by 40% in minutes

Access and Availability

Argon is currently only available through the Fairwind Program to cybersecurity partners — no public release date. Google is participating in the U.S. government's voluntary pre-release model access process, gathering feedback from early testers and iterating on guardrails before wider release.

This means most developers and enterprises can't access Argon directly in the short term — a stark contrast to OpenAI's GPT-6.1 Sol, announced the same day at DevDay 2026 and already available to Pro/Business users.

Developer Selection Guide

Sources

• Google Blog: https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/
• TechCrunch: https://techcrunch.com/2026/09/30/google-releases-gemini-4-argon-called-its-most-powerful-model-yet/
• MarkTechPost: https://www.marktechpost.com/2026/09/30/google-deepmind-unveils-gemini-4-argon-with-1m-output-tokens-for-coding-knowledge-work-and-cyber-defense/
• CNBC: https://www.cnbc.com/2026/09/30/google-gemini-4-argon-ai.html