Google launches Gemini 4 Argon at $2/$10 per million tokens

"Argon excels at deep work in complex, real-world tasks across domains including coding, finance, legal, cybersecurity, and science," wrote Koray Kavukcuoglu, SVP of Google DeepMind and Chief AI Architect, in Google's launch post on October 13, 2026.

Pricing lands at $2 per million input tokens and $10 per million output tokens, with cached inputs discounted 95%. Those are introductory rates. After Nov 30, 2026 the standard tariff is $4 and $20.

The benchmarks Google chose to publish are heavy on specialist domains. Argon ranks #1 on the Vals Index across finance, coding, legal and tax work, scores 87% on the ERQA embodied-reasoning benchmark, 57% on RareBench for rare-disease diagnosis, and 52.2% on LabBench. On the safety side, Google reports bio red-teaming vulnerabilities fell to "less than 1%", shortcut attempts on coding benchmarks to "just 2%", and attack success on the Gray Swan indirect prompt-injection test to 0.7%. All of these are self-reported in the launch post.

The engineering stories are specific. A Chrome team used Argon to pull core webpage handling out of a "~6 million line legacy folder" inside Chromium's "40 million-line codebase". Gemma inference on A100 saw a "2.85x end-to-end inference speedup", with throughput moving from roughly 37,900 to over 108,000 tokens per second. Internally, Google says agentic code submissions rose "by more than 35%" and code rollbacks dropped "by nearly a third."

Outside software, the post says Argon "nearly doubled" AlphaQubit inference throughput, cut spacetime costs for an arcsin implementation "by more than 100x", and that a video-to-sim-to-real robotics pipeline using the model landed "seven consecutive shots" in cornhole.

Access includes free top-tier Antigravity seats for 10,000 researchers for six months and an expansion of the Fairwind cybersecurity-defender program.

Kavukcuoglu closes: "Argon is just the beginning. We're looking forward to bringing you the rest of the Gemini 4 family."