Overview
- Google announced Gemini 4 Argon on Wednesday and presented it as a frontier AI built for complex, long-running technical tasks rather than general consumer chat.
- Argon raises its output limit to one million tokens, which Google says lets the model sustain much longer single-track reasoning and handle tasks like large code migrations and long video or document analysis.
- Google published benchmark scores that put Argon near the top on programming and security tests, citing results such as 77.9% on DeepSWE and a 68% tie on CWE-bench v1.
- In pilot work with security firm Wiz, Google reported Argon autonomously found and helped patch a critical vulnerability in widely used hospital software, but those vulnerability-discovery and repair claims are currently supported only by Google and its pilot partner, not broad independent verification.
- Google is beginning a controlled rollout through its Fairwind partner program and plans staged commercial access for Google AI Ultra subscribers and paid API customers at promotional token pricing, and the company says it built prompt-injection defenses and real-time monitoring though independent safety testing is still needed.