Google has launched Gemini 4 Argon, the first model in its Gemini 4 series, built with a focus on defensive cybersecurity. According to the company, the model can autonomously discover, validate, and patch software vulnerabilities. By spanning the full vulnerability lifecycle — discovery, validation, and patching — the announcement centers on remediation, the stage where security teams often accumulate backlogs of known weaknesses.

The launch was announced on September 30, 2026. While security is the headline use case, Google is also positioning Argon for long-running, complex workflows in software engineering and knowledge-intensive fields such as law and finance.

Benchmark performance

Argon's most prominent specification is output capacity: the model can generate up to 1 million tokens in a single response, a sharp increase over the 64K token output limit of earlier models. That capacity is aimed at large, continuous workloads, allowing an entire migration or review cycle to be produced in one pass rather than assembled across many smaller responses.

On DeepSWE v1.1, a benchmark built around real-world software engineering tasks, Argon scored 77.9%. That places it ahead of OpenAI's GPT-6 Astra at 74.1% and Anthropic's Claude Opus 5.5 at 74.2% on the same test, with all three frontier labs now measured side by side on real-world engineering work.

Security is where Google is emphasizing its lead. Argon recorded a top score of 68% on the vulnerability remediation portion of CWE-bench v1, a benchmark centered on fixing known categories of software weaknesses. CWE refers to the Common Weakness Enumeration, the standard taxonomy of software weakness categories maintained by MITRE, which gives the score a common reference point across vendors.

Google has also disclosed internal results. The model helped migrate more than800,000 lines of code in the Zircon kernel, part of the Fuchsia OS, over to Rust. That effort saved 300 TiB of memory across Google's data centers. Rust is a memory-safe systems language, and kernel rewrites of this kind are part of a broader industry effort to eliminate memory-safety vulnerabilities at the language level.

Availability and pricing

Argon is not yet broadly available. It is currently accessible through Google's Fairwind Program, which counts more than 650 vetted cyber defense organizations alongside internal Google teams. Participants include security firms CrowdStrike and Palo Alto Networks. Wiz, one of the early deployers, used the model to uncover significant flaws in widely used healthcare software. Until access widens, most developers will be judging Argon through results reported by Google and program participants, since the benchmark figures released so far are self-reported.

Wider availability is expected for paid API customers and subscribers to Google AI Ultra, though Google has not tied that expansion to a specific date in the material released so far. Because Argon is the first model in the Gemini 4 series, the timing of broader access and of follow-up releases in the family are the immediate items to watch.

On pricing, Argon starts at an introductory rate of $2 per million input tokens and $10 per million output tokens. Under the published roadmap, those rates can rise to $4 and $20 respectively, a defined escalation path that gives prospective customers the introductory window to evaluate the model before full rates apply.