Google launches Gemini 4 Argon, a frontier AI model aimed at coding and cybersecurity

Google introduced Gemini 4 Argon on September 30, 2026, describing it as the newest entry in its Gemini line of artificial intelligence models and its most capable system to date for complex, multi-step professional work, according to Google's own announcement on its blog.

The company says Argon is built for "long-horizon" tasks that require sustained reasoning, including software engineering, enterprise knowledge work such as legal drafting and financial research, and cybersecurity defense, including finding and patching software vulnerabilities.

What Argon can do

According to Google, Argon supports an output limit of up to 1 million tokens, which the company describes as industry-leading and a large increase from the 64,000-token limit of earlier models. Google also highlights the model's visual understanding, including analysis of documents and video, and its ability to work through multi-step coding and vulnerability-patching tasks autonomously.

On benchmark tests Google cited in its announcement, Argon scored 77.9% on DeepSWE v1.1, a test of real-world software engineering tasks, and 51.3% on AutomationBench, where Google says it ranks first. The model also tied for first place on CWE-bench v1, a benchmark for fixing security flaws, with a score of 68%, and led a prompt-injection robustness benchmark run by Gray Swan, per Google's blog.

TechCrunch reported that Google compared Argon's benchmark results to recent releases from competitors, including OpenAI's GPT-6 Astra model and Anthropic's Fable and Opus models, saying Argon scored higher on the tests Google selected. TechCrunch also reported that Google pointed to its Gemini app surpassing one billion monthly users, a figure the company says now matches ChatGPT's user base.

Limited rollout and pricing

Argon is not yet broadly available. Google said access is starting with "trusted cyber defenders" through its Fairwind Program, a security initiative that TechCrunch described as a select group of the company's cyber partners. Google said broader access for developers, enterprises and consumers would follow "as soon as possible" but did not give a firm date.

For API access, Google listed an introductory price of $2 per million input tokens and $10 per million output tokens, with a stated 95% discount for cached inputs. The company said standard pricing of $4 and $20 per million tokens, respectively, would apply later.

Google's announcement also emphasized safety measures built into the model's deployment, including defenses against prompt-injection attacks, misalignment monitoring, and sandboxed testing environments used before any wider release.

Internal codenames: Argon, Barium and Carbon

Reporting by Business Insider, cited and summarized by the technology outlet TestingCatalog, indicates that Google tested several internal checkpoints of Gemini 4 before choosing one to release publicly. According to that reporting, Google's internal "Barium-B" checkpoint is the version that was released publicly as Gemini 4 Argon.

The same reporting says Google employees have since been testing a further, separate checkpoint internally named Carbon through the company's internal development environment, referred to as Jetski. Based on early, unnamed employee feedback cited in Business Insider's reporting, Carbon's coding performance was described as comparable to Anthropic's Claude Opus 5.5, though the feedback cautioned that more testing was needed. TestingCatalog reported that it remains unclear whether Carbon will become a future update to Argon or a separate release, and that Google has not commented on its internal roadmap.

A separate workplace AI agent

On October 8, 2026, at an event called Gemini at Work, Google Cloud separately announced a product it calls the Gemini agent, described in Google's own announcement as a "universal agent for work." According to that announcement, the agent plans the work, uses skills and tools, and connects to a company's business systems from a single prompt, drawing on multiple underlying AI models and choosing what Google describes as the best model for each job.

Google's post says the agent delivers finished results inside the documents, inbox and developer environments that customers already use. Google's announcement of the agent does not specify which underlying Gemini model or models power it.

What is not yet known

• Google has not given a specific date for when Argon will be available to the general public or to all paid developers.
• It is not confirmed whether the internally tested "Carbon" checkpoint will ship as a future Argon update or as a distinct release.
• Google's own blog post on the Gemini agent for work does not state which Gemini model version underlies that product.