The Pulse
Google Opens Gemini 4 Argon to Cybersecurity Defenders
Google announced Gemini 4 Argon on September 30, positioning the model for software engineering, legal and financial work, and cybersecurity defense. The company is first giving access to trusted cyber defenders and says broader access will
AI.info Team · Written with AI ·
Google announced Gemini 4 Argon on September 30, introducing a flagship model aimed at lengthy software-engineering, legal, financial and cybersecurity tasks. The company is not opening general access yet: its first rollout goes to trusted cyber defenders through the Fairwind Program, while Google gathers feedback and adjusts safeguards.
“Safely releasing frontier capabilities at this level requires a phased approach,” writes Koray Kavukcuoglu, SVP of Google DeepMind and Google’s chief AI architect. Google says the model will reach developers, businesses and consumers later, starting with paid API customers and Google AI Ultra subscribers, but gave no date for that expansion.
Argon stretches output to one million tokens
Google says Argon supports up to 1 million output tokens, up from a previous limit of 64,000. That larger ceiling is designed for tasks where a model needs to work through many steps and produce extended responses, rather than hand back a short answer after each stage.
The company’s published results show mixed performance across the evaluations it lists. Argon scores 77.9% on DeepSWE v1.1, which Google describes as a test of long-horizon software-engineering work, and 51.3% on AutomationBench, a benchmark for end-to-end business execution. On Harvey’s Legal Agent Benchmark, Google reports a score of 19.6%; results on that evaluation are not a direct measure of accuracy in every legal workflow.
Google also reports 91.7% on LVBench, which tests long-video understanding, and says Argon ties for first on CWE-bench v1 with a score of 68% for security-vulnerability remediation. Those figures are company-published benchmark results, not evidence on their own of how the model will perform in customers’ production systems.
Cyber defense shapes the first release
Google says Argon can find, validate and patch software vulnerabilities autonomously. The staged launch reflects the dual-use risks of a model with those capabilities: the company says it will first provide access to trusted defenders and its own internal teams, while continuing work on safeguards before broader release.
Google says it is participating in the U.S. government’s voluntary pre-release model-access process. It also says the model’s safeguards are being tested against harmful cyber and chemical, biological, radiological and nuclear requests, indirect prompt-injection attacks, and misaligned actions. Those protections are still being strengthened before public availability.
The company cites early work with cybersecurity firm Wiz through its Scan for Good initiative. Google says Argon identified a critical vulnerability that exposed sensitive personal information in healthcare software used by hospitals worldwide; the announcement does not name the software or provide details that would allow independent assessment of the finding.
Google’s internal examples come with a rollout caveat
Google says thousands of its employees already use Argon for specialized coding, research and writing. In one internal example, the company says agents rewrote 32,000 lines of SIMD code in libgav1, Google’s video decoder, and produced a memory-safe Rust version that ran 2.7 times faster than an existing Rust port while returning identical video output.
That example remains Google’s account of its own internal work. External developers cannot yet test Argon for themselves, and Google has not set a public release date. The introductory API price is $2 per million input tokens and $10 per million output tokens; after the introductory period, the listed rates rise to $4 and $20, respectively. Cached input tokens receive a 95% discount during the introductory period.