Google has moved Gemini 4 Argon from its trusted-partner Fairwind program to the general public, opening the flagship model to AI Pro and Ultra subscribers across the Gemini app, AI Mode in Search and Antigravity as of yesterday’s global rollout. Developers get it too: the paid Gemini API tier, Google AI Studio, Gemini Enterprise plans and OpenRouter (as google/gemini-4-argon) are all live.
Pricing is the headline for builders. The introductory rate runs $2 input / $10 output per million tokens through November 30, with cached inputs at 95% off; list prices double to $4 / $20 on December 1. That slots Argon at parity with Claude Sonnet 5.5’s intro pricing and well under GPT-6 Astra’s $10 / $50.
What Google claims – and what holds up
Google’s vendor benchmarks put Argon at 77.9% on DeepSWE v1.1 (vs Claude Opus 5.5’s 74.2% and GPT-6 Astra’s 74.1%), 68% on CWE-Bench, 91.7% on LVBench, and a 15% hallucination rate on AA-Omniscience. Independent trackers are kinder than most launch-day hype: Vals’ GDP-weighted index (Oct 12) ranks Argon #1 at 68.9%, while Artificial Analysis ties it with GPT-6 Astra at 53 and still scores Claude Opus 5.5 and Sonnet 5.5 higher overall. Translation: genuinely strong, vendor slides admittedly vendor-y.
One million tokens of output, minus the asterisks
Argon ships a unified one-million-token input/output limit via a continuation protocol, up from 64,000 tokens of output – though Google hasn’t clarified the maximum uninterrupted response size, so long-form agents should plan around checkpoints. Terminal-Bench 4.0 is the notable loss for Google (57.4% vs Opus 5.5’s 66.4%), while Harvey’s legal-agent benchmark is a blowout (19.6% vs GPT-6 Astra’s 5.4%).
For crypto and on-chain builders, the practical read: frontier-class agentic coding just got a cheaper hourly rate for six weeks, and the December 1 price double is the deadline to stress-test whatever you build on it.
