Google launches Gemini 4 Argon, its most advanced AI model yet

Google has unveiled Gemini 4 Argon, its most advanced AI model yet, pushing deeper into coding, cybersecurity, and professional work as the race for frontier AI shifts from chatbots to systems that can carry out long, multi-step jobs.

Argon won’t have a typical public launch. Google is initially giving access to a limited group of trusted cyber defenders through its Fairwind Program, with wider availability planned after further safety testing and pre-release evaluations with the U.S. government.

Announcing the launch on X, Alphabet CEO Sundar Pichai framed Argon as an early look at Google’s next major model.

“Lots of discussion out there about our next model(!), so I wanted to give an early look as soon as possible. Introducing Gemini 4 Argon! It shows frontier performance in complex workflows, cyber defense and software engineering. Teams are using it extensively at Google, from coding to quantum computing, great feedback.”

“Starting this rollout in this way gives us more confidence, but also enables us to put a model that is trained and strong in cyber defense in the hands of defenders as soon as possible,” Tulsee Doshi, Google’s Gemini model product lead, told CNBC.

The phased release reflects what Google says Argon can do. The model can autonomously search for software vulnerabilities, validate them, and generate patches. On CWE-bench v1, a test focused on fixing security flaws, Argon scored 68%, tying for first place. Google says it produced a large jump in vulnerability discovery compared with Gemini 3.8 Flash Cyber, which arrived in September.

Gemini 4 Argon Performance on Coding Benchmark

Argon is aimed far beyond cybersecurity. Google says the model scored 77.9% on DeepSWE v1.1, setting a new high on a benchmark that measures performance on long-horizon, real-world software engineering tasks.

Google engineers are already using Argon for debugging, large codebase migrations, and algorithm design. The model also ranked first on the Vals Index, which measures AI performance across economically significant work in finance, coding, legal, and tax, and scored 51.3% on Zapier’s AutomationBench for end-to-end business workflows.

Argon also posted strong results on visual knowledge tasks, including a 91.7% score on LVBench, a benchmark for long-video understanding.

Google is already putting Argon to work inside its own infrastructure

Some of Argon’s most interesting results are coming from Google itself.

Google says teams used groups of Argon agents to analyze data-center telemetry and identify memory optimizations that freed more than 300 tebibytes of memory after deployment. The company estimates the work could eventually save between 500 TiB and 1 PiB of memory capacity. That matters at a time when AI companies are spending heavily on servers, chips, networking equipment, and data centers to keep up with soaring compute demand.

Google’s quantum computing researchers have tested Argon on algorithm optimization. In one example the company cited, the model improved a published baseline by 40% in minutes. Engineering teams are using Argon for large code migrations, including efforts to move C and C++ software into Rust.

Google is giving the model far more room to work on those kinds of jobs. Argon supports output of up to 1 million tokens, up from 64,000 tokens in earlier systems, allowing a single run to continue through far longer reasoning and execution sequences. Google plans introductory API pricing of $2 per million input tokens and $10 per million output tokens, later rising to $4 and $20.

The release comes one day after Google CEO Sundar Pichai joined President Donald Trump and executives from several major AI companies in signing a voluntary White House agreement centered on safeguards for increasingly capable AI systems.

Google says it is strengthening protections around model misuse, prompt injection attacks, agent behavior, and secure testing environments before Argon reaches a wider audience. Trusted cyber defenders will receive broader access to its security capabilities during the early rollout, giving researchers a chance to test the model against real systems before Google puts it into the hands of developers, enterprises, and consumers.

“We really believe that a model of this caliber and this level of frontier performance is meaningfully important for defenders,” Doshi said.

That staged release may prove to be the more consequential part of the announcement. The race among Google, OpenAI, Anthropic, xAI, and other frontier labs is moving past which model writes the best answer. Increasingly, the question is which systems can be trusted to act, code, research, and operate across real infrastructure with far less human supervision.