Google Introduces Gemini 4 Argon as AI Firms Develop Personal Agents

Google’s Gemini 4 Argon is entering testing with trusted cybersecurity defenders as technology companies compete to develop personal agents intended to handle tasks across apps and services.

Google announced the model Sept. 30 for software engineering, business research and cybersecurity defense. Its initial rollout is through the Fairwind program, with broader access to developers, businesses and consumers planned as testing continues. No public release date has been announced.

Argon posted mixed results on coding benchmarks. It scored 77.9% on DeepSWE v1.1, compared with 74.1% for GPT-6 Astra, but scored 55.0% on FrontierSWE v2, below Astra’s 65.5%. On the CWE-bench v1 cybersecurity benchmark, the models tied at 68%.

The benchmark results do not establish how either model will perform across everyday tasks. Google is gathering feedback from early testers and refining safeguards before making Argon more widely available.

Google introduced Gemini Spark in May as a planned personal agent. The company plans to offer it through the Gemini app, with connections to Google services and other tools over time. Initial access was planned for trusted testers and Google AI Ultra subscribers in the United States.

Meta introduced Muse in the United States on Sept. 8, describing features such as sending email and booking travel. Meta said the agent would be free for most uses, with paid plans for heavier use. OpenAI introduced its Dots agents on Sept. 29, describing agents that can work continuously, use their own cloud computers and connect to more than 4,000 apps through plugins.

The announcements outline planned capabilities, but do not establish how reliably these agents complete everyday tasks. Google has said it will continue testing Argon and refining its safeguards before broader access; it has not set a public release date.