Google Announces Gemini 4 Argon With 1M Output Tokens and Deep Reasoning
Google has introduced Gemini 4 Argon, a frontier AI model featuring a 1-million-token output limit, advanced reasoning for complex code migrations, and strong enterprise benchmarks.
Google has unveiled Gemini 4 Argon, a frontier AI model engineered to handle deep reasoning across multi-step, long-horizon tasks in software engineering, enterprise operations, and scientific research. The company is taking a phased deployment approach, initially providing early access to vetted security professionals through Google’s Fairwind Program while engaging in the U.S. government’s voluntary pre-release evaluation process. Feedback gathered from early testers will guide guardrail refinements before broader availability opens to developers, businesses, and consumers.
The model is slated to launch at an introductory rate of $2 per million input tokens and $10 per million output tokens, alongside a 95% discount for cached inputs. A major architectural upgrade in Argon is its 1-million-token output limit—a notable leap from previous 64,000-token limits—giving the system sufficient room to perform complex reasoning trajectories in a single run.
Proven Results Across Internal Systems
Ahead of external availability, Google deployed Argon across internal engineering environments to optimize systems, conduct technical research, and refactor legacy codebases. In cloud infrastructure, autonomous Argon agents evaluated fleet telemetry to discover and apply memory optimizations across Google data centers. The intervention reclaimed over 300 TiB of memory upon deployment, with expected total savings between 500 TiB and 1 PiB.
Argon has also assisted with major language migrations, converting critical C/C++ libraries and services to safe Rust. Target projects include foundational components like re2, the libgav1 video decoder, and the Fuchsia Zircon kernel, which contains more than 800,000 lines of code. For libgav1, Argon agents analyzed an existing Rust port, ran iterative profile-guided experiments, and swapped out 32,000 lines of SIMD instructions with compiler-vectorized safe Rust. The resulting decoder maintains identical video output while running 2.7 times faster than the previous Rust port, closing the performance gap with C++. In quantum computing operations, the model helped optimize space-time resources for algorithmic subroutines, beating published benchmarks by 40% in minutes.
Benchmark Records, Cybersecurity, and Science
Across external evaluations, Argon demonstrated state-of-the-art results in long-horizon software engineering, securing a 77.9% score on DeepSWE v1.1. It also achieved top marks on Zapier’s AutomationBench (51.3%), LVBench for long video comprehension (91.7%), and the Vals Index, which assesses economic tasks across legal, finance, coding, and tax domains.
Defensive cybersecurity represents a central application for Argon. On CWE-bench v1, which evaluates vulnerability remediation, Argon tied for the top spot at 68%. Security firm Wiz integrated Argon into its Scan for Good initiative, using the model to identify a high-risk data exposure flaw in healthcare software that earlier frontier systems missed. To assist defensive teams, Google plans to offer unguardrailed access specifically for trusted defenders and internal security units. Public releases will incorporate robust protections, including defenses against indirect prompt injections where Argon leads Gray Swan’s IPI benchmark, runtime monitoring of reasoning traces, and isolated sandbox environments.
Scientific deployments have yielded similar momentum. Researchers at Gamow Labs used Argon to analyze newborn genomic data, discovering an overlooked chromosomal rearrangement linked to rare lung conditions and leading both RareBench (57%) and LabBench (52.2%). In fusion energy research, a Google team collaborating with Professor Matt Landreman discovered new counterexamples to Grad’s conjecture regarding stellarator magnetic fields. The model has also generated functional serine esterase enzyme designs within hours and scored 87% on the Embodied Reasoning (ERQA) benchmark, later generating simulation code that enabled Apptronik’s Apollo humanoid robot to land seven consecutive cornhole shots.
What it means for developers
Gemini 4 Argon’s technical profile indicates a growing industry emphasis on autonomous, long-context development tasks. The combination of an industry-leading 1-million unified input-output token capacity and a continuation protocol allows developers to assign extensive multi-step workflows—such as full-repository migrations, comprehensive security audits, and multi-file code synthesis—without running into tight generation constraints.
Developers can access the model via the Gemini API paid tiers within Google AI Studio and Android Studio, as well as through Google Antigravity for Google AI Ultra subscribers. The combination of aggressive input caching discounts and massive output windows reduces financial and technical hurdles for building reasoning agents. As frontier options expand across different providers, developers can try top AI models cheaply through one API at https://apixoai.online to evaluate alternative architectures side by side. For technical teams, Argon highlights how frontier AI is shifting away from simple prompt-response interactions toward persistent agents that handle end-to-end software delivery.
Source: Google Drops Gemini 4 Argon: Everything You Need to Know — Towards AI. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
