Google released Gemini 4 Neon on Wednesday, the fourth model in its Gemini 4 family in under seven weeks, positioning it between Carbon for the hardest long-running projects and Hydrogen for fast everyday work. The company pitches Neon as a model for complex work that gets repeated — software engineering, analytical reasoning and business workflows — with fast end-to-end task completion across coding, knowledge work and science.
Access is staged across the stack: developers can call it through the paid Gemini API in Google AI Studio and Android Studio; businesses get it via the Gemini for business API and model selector; paid Google AI Pro and Ultra subscribers can select it in Antigravity and the Gemini app. Google says eligible Fairwind Program defenders will receive a Neon variant tuned for cybersecurity within days.
Neon carries a 1 million-token context window that accepts text, images, video, audio and PDFs, but does not generate images or audio, and the Live API is not supported. Pricing is $3 per million input tokens and $15 per million output tokens, with cached input at $0.15 — a 95 percent discount. That makes it twice the price of Hydrogen at $1.50/$7.50, but 25 percent cheaper than Carbon's standard rates, which rise from an introductory $2/$10 to $4/$20 on December 1.
Benchmark claims are company-reported and configuration-dependent. At its xhigh thinking setting, Google cites a 61 on the Artificial Analysis Intelligence Index, third overall behind larger, slower models, and a first-place 58.6 on the Artificial Analysis Coding Agent Index. On the Vals Index it ranks second at 70.9 percent, and on Arena's new Agent Arena Index it debuts second worldwide at 1,336 across more than 4,200 real-world agent sessions, behind Anthropic's Claude Fable 5.2.
Efficiency is the sharper claim: at the default high setting, Google says Neon costs 32 percent less per task than Carbon on the Artificial Analysis intelligence evaluation, and roughly 50 to 80 percent less than top competing frontier models on the Vals and Agent Arena evaluations. Third-party Arena data pegs the median cost at $2.82 per task — 53 percent cheaper than the nearest competitor, Claude Opus 5.5. Those are benchmark comparisons, not price cuts to the listed rates.
Coding results show how much the surrounding harness matters: Neon scores 12.8 percent on SWE-sweep under the mini-SWE harness but 39 percent inside Google Antigravity. It also posts 65.2 percent on Terminal-Bench 4.1, first place on Zapier's AutomationBench at 56.3 percent, first on Andon Labs' Vending-Bench 2 at $16,843 in a simulated long-horizon business run, and Google reports its lowest fabrication rate yet in internal Antigravity testing.
Four thinking levels — low, medium, high and xhigh — let teams trade token spend for reasoning depth, and Google recommends choosing the lowest setting that meets the quality bar. The headline scores use xhigh, so teams benchmarking Neon against competitors should budget for the additional reasoning tokens.
The release continues Google's aggressive family cadence — Argon in late September, Carbon in late October, Hydrogen on November 12, now Neon — and reflects a market shifting from "which model is smartest" to "which model finishes the work fastest per dollar." A model that lands third on intelligence while leading coding-agent rankings at a fraction of flagship cost is aimed squarely at production budgets, where most token spend now lives.
Comments (0)
Log in to join the discussion
Log InNo comments yet