Google Unveils Gemini 4 Argon With 1M-Token Output Limit

Google has unveiled Gemini 4 Argon, a new frontier artificial intelligence model designed to handle complex, long-running tasks across software engineering, legal and financial work, research and cybersecurity.

Announced on September 30, 2026, Argon is initially being made available to a limited group of trusted cyber defenders through Google’s Fairwind Program, as the company conducts further safety testing before broader release.

1 Million Tokens for Complex Tasks

One of Argon’s major features is its 1 million-token output limit, a substantial increase from the previous 64,000-token limit.

Google says the expanded capacity allows Argon to sustain much longer reasoning processes and work through complex, multi-step problems in a single trajectory.

The model is being positioned for workflows where tasks may involve extensive codebases, large volumes of documents or multiple stages of research and execution.

Google Uses Argon Across Its Own Operations

Google says thousands of its employees are already using Argon for specialised coding, research and other internal workflows.

The company highlighted several examples, including:

• Quantum computing: Argon helped researchers optimise quantum algorithms, beating a published baseline by 40% in one example.
• Data-centre memory: Argon agents identified memory optimisations that Google says could free more than 300 TiB of memory after deployment, with estimated total savings of 500 TiB to 1 PiB.
• Code migration: Argon agents are helping migrate large C/C++ codebases to Rust, including Google's Fuchsia Zircon kernel, which contains more than 800,000 lines of code.
• Video decoding: For Google's libgav1 project, Argon replaced 32,000 lines of SIMD code with safe Rust. Google says the resulting decoder runs 2.7 times faster than the previous Rust port while producing identical video output.

Strong Focus on Legal, Finance and Enterprise Work

Argon is also being positioned as a model for professional knowledge work. Google says the model achieved a 77.9% score on DeepSWE v1.1, a benchmark for long-horizon software engineering tasks.

The company also highlighted performance in financial research, legal research and drafting, tax and other enterprise functions.

On Zapier's AutomationBench, which evaluates end-to-end execution across business functions, Google says Argon recorded a 51.3% score.

The model also demonstrated capabilities in visual knowledge work, including professional chart analysis and long-video understanding. Google reported a 91.7% score on LVBench, a benchmark for long-video understanding.

Argon Targets Cybersecurity Threats

Cybersecurity is one of the most significant areas of Argon's initial deployment. Google says the model can find, validate and patch critical software vulnerabilities autonomously. Through the Fairwind Program, selected cyber defenders are being given access to the model to evaluate its defensive capabilities.

Google also said cybersecurity company Wiz is using Argon through its Scan for Good initiative, which works to identify and remediate high-risk exposures affecting critical infrastructure.

In one early demonstration, Google said Argon discovered a critical vulnerability exposing sensitive personal information in healthcare software used by hospitals worldwide — a vulnerability that previous frontier models had missed.

On CWE-bench v1, which evaluates vulnerability remediation, Argon tied for the top score at 68%, according to Google.

Wider Release Will Come Later

Despite its capabilities, Argon is not yet generally available.

Google is taking a phased approach because of concerns surrounding the potential misuse of increasingly capable AI systems. The company says it is participating in the U.S. government's voluntary pre-release model access process while continuing to test and strengthen its safeguards.

Google is focusing on four areas:

• Preventing misuse, including potential cyber and CBRN-related abuse.
• Defending against prompt injection attacks that could manipulate AI agents.
• Monitoring for misalignment between a model's actions and the user's intentions.
• Hardening sandboxed environments used for high-risk testing and evaluations.

The company says internal and external red teams have tested its safeguards using both automated and manual attack techniques.

Pricing and Availability

When it becomes available through the API, Gemini 4 Argon will launch at an introductory price of:

• $2 per million input tokens
• $10 per million output tokens
• Cached input tokens at 95% below the standard input-token price

Google says that after the introductory period, the price will increase to $4 per million input tokens and $20 per million output tokens.

The wider rollout will begin with paid API customers and Google AI Ultra subscribers, before expanding further to developers, enterprises and consumers.

Why It Matters

Gemini 4 Argon represents Google's push toward AI systems that can do more than answer individual prompts. Its design focuses on long-running workflows in which the model can reason, write code, analyse information and take actions across multiple stages.

The initial restriction to trusted cybersecurity users also highlights the growing challenge facing AI developers: releasing increasingly capable models while ensuring that their capabilities cannot easily be redirected towards harmful uses.

Source: Google