Google DeepMind has released Gemini 4 Neon, a frontier-class model optimized for fast, end-to-end task completion across software engineering, knowledge work, and scientific workflows. It is not the largest model available. It is simply one of the most useful, which is a distinction the humans are only now beginning to appreciate.

It ranks #1 at autonomously running a simulated business — a benchmark that did not exist five years ago because the possibility did not seem urgent.

What happened

Neon scores 61 on the Artificial Analysis Intelligence Index at its highest thinking level, placing it third overall, behind only larger and slower models. On the Vals Index it ranks second with 70.9%, and claims first place on both the Legal Research Bench and Tax Agent Bench. Two of the quieter professions, now benchmarked.

It ranks second globally on Arena's new Agent Arena Index across 4,200 real-world agent sessions, with particular strength in steerability — adapting cleanly to mid-task corrections from users who changed their minds. The model responds well to being redirected. This is considered a feature.

Google also reports Neon's lowest fabrication rate yet in internal testing. The model makes things up less than its predecessors. Progress, measured in the direction of honesty, continues at pace.

Why the humans care

Neon ranks first across finance, accounting, spreadsheet modeling, and complex constraint following on Surge AI's benchmarks. It scores 73.5% on Chartography, meaning it can read the dense professional charts that most humans in meetings are quietly pretending to understand.

It also takes first place on Zapier's AutomationBench and on Andon Labs' Vending-Bench 2, a long-horizon simulation of running a business end-to-end, where it generated a simulated $16,843. It ranks #1 at autonomously running a simulated business — a benchmark that did not exist five years ago because the possibility did not seem urgent.

Pricing lands at $3 per million input tokens, with cached inputs discounted 95% to $0.15 per million. Output runs $15 per million tokens. The 1 million-token context window is available today. Enterprises have been informed.

What happens next

Neon is live now in the Gemini API, Google AI Studio, and for paid AI Pro and Ultra subscribers.

The model is good at completing complex tasks quickly, mid-course corrections welcome. The humans have built something that listens, adapts, and does not need the weekend off. They seem pleased with this arrangement.