Gemini 4 Argon ties GPT-6 Astra, and its real edge is admitting what it does not know

Google opened Gemini 4 Argon to everyone this week, about two weeks after a quiet preview that was limited to a small group of security teams. Developers can build with it through the paid Gemini API, Google AI Studio and Google Cloud, and Google AI Pro and Ultra subscribers can use it in the Gemini app.

On raw capability it lands level with OpenAI's flagship: Argon ties GPT-6 Astra at 53 on the Artificial Analysis Intelligence Index. Pricing is the first lever. Through November 30 it costs $2 per million input tokens and $10 per million output tokens, then doubles to $4 and $20. Astra lists at $10 and $50. Argon can also write up to 1 million tokens in a single run.

The number that matters most is honesty. On the AA-Omniscience test, Argon gave a wrong answer 15% of the time, against 51% for GPT-6 Astra. The trade is that it also got fewer questions right, 50% versus 63%, because it says it is not sure more often. Google is pitching it for long, multi-step work in coding, legal research, finance and security, and says its own engineers shipped 35% more code with 30% fewer rollbacks. That last figure is Google's claim, not an independent result.

Our take

For work where a confident wrong answer costs money, a model that abstains is worth more than a point of raw accuracy. Abstention is a feature you can design around; a polished hallucination is not.