You ask an AI to compare documents, then learn mid-task that only the signed versions count. A fast answer doesn’t help if it ignores the correction or forgets the deadline you already gave it. That’s the interesting pitch behind Gemini 4 Neon, announced on November 18: adjust course on a multi-step job instead of treating every new message as a request to start over.
What the ranking actually measures

Google highlights Neon’s second-place steering result on Agent Arena. That’s the ability to incorporate corrections during a task. It doesn’t make it second at everything or guarantee your interface lets you intervene at every stage. Google also showed Neon coordinating agents in Antigravity to recreate Alan Turing’s study in interactive 3D. That’s a long-running task, not an independent steering test.
Where to find it and what it costs
Google began rolling out Neon via the Gemini API, Google AI Studio and Gemini for business, and to eligible Pro and Ultra subscribers in the Gemini app and Antigravity. The API costs $3 per million input tokens, $0.15 per million cached input tokens and $15 per million output tokens. The context window is one million tokens. That’s room for more material, not an assurance that every detail will get noticed. API pricing and subscription pricing are also separate.
Mais FPS take
Give two models the same task and three concrete rules. Partway through, change one. Check whether the result reflects the correction, preserves the other two rules and shows what got revised. Use test documents and avoid irreversible actions. If changing one detail still means starting over, the time savings haven’t reached your workflow.
Sources
Google — Gemini 4 Neon · Agent Arena
Cover: official image from Google.
