Google unveils Gemini 4 Argon, tops OpenAI and Claude

Google's own benchmark disclosures show a model that wins broadly but not everywhere. Argon tops Vals Index (68.9%), Vals Finance Agent v2 (65.4%), and AutomationBench (51.3%), yet loses to GPT-6 Astra on FrontierSWE v2 (55.0% versus 65.5%) and to Claude Sonnet 5.5 on Terminal Bench 4. Pichai framed the launch as an early look, prompted by talk of Google's next model rather than by a fixed release calendar.

Beyond scores, Google points to internal deployment as proof: quantum algorithm optimization improved compute use by 40% against an existing research baseline, data center agents cut memory use by more than 300TiB with 500TiB to 1PiB expected, and Rust migration work on the libgav1 video decoder sped up 2.7x while keeping identical video output. Those are unusual disclosures that mix research claims with operational results.

The cybersecurity posture is the sharpest constraint. Because Argon can find, verify, and fix vulnerabilities autonomously, Google is gating access through Fairwind, participating in a US government pre-release review process, and citing a 0.7% attack success rate on Gray Swan's indirect prompt-injection benchmark. The outcome hinges on whether that caution holds as access widens to paid API and Google AI Ultra users, and on whether buyers trust Google's self-reported numbers over third-party tests.