Google has unveiled Gemini 4 Argon, a frontier AI model with a 1 million-token output limit — 16 times the 64,000-token ceiling of previous generations.

Announced on September 30, Argon targets software engineering, knowledge work and cybersecurity. The expanded output lets the model work through long, multi-step tasks in a single run, such as migrating an entire codebase or drafting a long legal document, without losing its planning state between calls.

Google says Argon leads or ties GPT-6 Astra in 14 of 19 published benchmarks, though Astra scores higher on some coding tests. Internally, Argon agents freed more than 300 terabytes of data-centre memory and rewrote 32,000 lines of video-decoder code to run 2.7 times faster.

Argon is rolling out first to vetted cyber defenders through Google’s Fairwind programme, with paid API tiers and AI Ultra subscribers next. Free Gemini users get no access yet, and Google has given no date for wider rollout.

The launch sharpens the frontier-model race with OpenAI’s GPT-6 Astra and Anthropic’s Claude Opus 5.5.