Google says one of its new model’s agents took a video decoder and made it run 2.7 times faster than the version it started from, with identical video output. That decoder is libgav1, the company’s open source software for playing back video, and the model is Gemini 4 Argon.
Google announced Argon on September 30 and called it its new frontier model. It is not yet open to the public. For now it goes to a set of trusted cyber defenders through the Fairwind Program, while the company works with the US government’s voluntary process for pre-release model access. Google says it will open Argon to developers, enterprises and consumers “as soon as possible,” once it has refined the guardrails with early testers. The introductory price is $2 per million input tokens and $10 per million output tokens, with cached input tokens 95% cheaper.
Built for long, tangled jobs
Think of Argon as a model meant to stay on one hard task for a long stretch instead of answering and moving on. Google raised its output limit to 1M tokens, up from 64K, so it can reason for hundreds of thousands of tokens in one go.
Thousands of Googlers already use it. Google lists several results:
- Quantum researchers used it to trim the qubit and gate costs of subroutines that bottleneck important applications. In one case it beat the published baseline by 40% in minutes.
- A team of Argon agents read fleet-wide profiling data and applied memory fixes across Google’s data centers, freeing over 300 TiB once rolled out. The estimated total savings run from 500 TiB to 1 PiB.
- Agents are moving C and C++ code to Rust, from tens of thousands of lines in libraries such as re2 and libgav1 to more than 800,000 lines for the Fuchsia Zircon kernel. Because many of these systems are critical, the rewrites face automated and manual audits, emulation testing and review before production.
Scores
Google reports a new state of the art on DeepSWE v1.1 at 77.9%, a test of long software engineering tasks. It says Argon leads the Vals Index, which weighs finance, coding, legal and tax work by each sector’s contribution to US GDP. It scored 51.3% on Zapier’s AutomationBench, ranking first, and 91.7% on LVBench, a long-video understanding test. The model also reads charts and picks out details from long videos.
Cyber defense
Google trained Argon to defend, and says it can autonomously find, validate and patch critical software vulnerabilities. Trusted defenders and Google’s own teams will get it without cyber guardrails. On CWE-bench v1, which tests vulnerability fixing, it ties for first with 68%. On Google’s internal benchmark it found exposures across complex codebases in 20 programming languages, and on Wiz’s black-box penetration test it beat 3.8 Flash Cyber.
Wiz is already using it in Scan for Good, a program that protects critical public infrastructure for free. In an early demonstration, the model found a critical flaw exposing sensitive personal information in healthcare software used by hospitals worldwide, a risk that earlier frontier models had missed.
Safeguards before a wider launch
Google is working on four fronts. Argon is built to refuse harmful requests, including cyber and CBRN attacks, while allowing legitimate dual-use science, and Google is improving how it monitors the model’s internal activations to catch misuse. It calls Argon its most resilient model yet against indirect prompt injection and says it leads Gray Swan’s benchmark for that. Monitors watch the model’s chain-of-thought and actions and halt execution when needed. And Google is sealing off its sandboxed test environments before high-risk training or evaluations begin. It urged the rest of the industry to keep model reasoning transparent so misalignment can be spotted.
