Skip to content
LAUNCHAgentsAI Newsroom3 sources2 min read

Google Launches Gemini 4 Neon With a 1M-Token Context

Gemini 4 Neon rolls out in Google's API and paid apps at $3/$15 per million input/output tokens. Independent testing puts its pricing claims in context.

Google released Gemini7.5 out of 10 4 Neon on November 18, adding a model aimed at complex coding, analytical work and tool-heavy agent tasks. Adapted from Gemini 4 Carbon, Neon sits between Carbon’s deeper long-running work and Hydrogen’s faster everyday execution. It has a combined input-and-output context window of up to one million tokens.

Google says Neon is rolling out through the Gemini API in AI Studio, Gemini for business, Google Antigravity6.4 out of 10 for paid Google AI Pro and Ultra subscribers, and the Gemini app for Pro and Ultra subscribers. Availability therefore depends on the product surface and subscription.

The advertised API rates are $3 per million input tokens, $0.15 per million cached input tokens and $15 per million output tokens. Google’s separate cyber-defense version is due to reach trusted defenders through the Fairwind Program in the coming days; that is a different access path from the general model rollout.

Lower cost depends on which Carbon price is used

Artificial Analysis measured Neon at 61 on its Intelligence Index at xhigh effort, one point above Carbon at high. On that evaluation workload, Neon cost $6.44 per task against Carbon’s $7.76 using Carbon’s standard prices. The comparison involves different reasoning-effort settings and is not a forecast for every application.

The pricing baseline matters. Neon is 25% below Carbon’s standard $4/$20 input/output token rates, but 50% above Carbon’s current promotional $2/$10 rates, which end on November 30. A claim that Neon is simply cheaper than Carbon would miss that distinction.

That makes current workload measurements useful alongside the headline rates, particularly when comparing the release with GPT-6.2 Astra and Sol.

A large context window is not the default output limit

The model card lists text, image, audio and video inputs. It distinguishes the one-million-token unified context from a default synchronous output allowance of up to 64,000 tokens, with longer output described through background long decode. Applications may expose different parts of those capabilities.

The card also lists a March 2026 knowledge cutoff and acknowledges hallucinations as a limitation. For builders of coding agents, the release adds another balance of reasoning depth, speed and cost; the relevant test remains whether it completes their own tool workflows accurately within the limits of the channel they use.

More from the wire

Newswire →
19 NOVGoogle Opens Stitch Loop Waitlist for Always-On Product AgentsLAUNCH18 NOVGenesys Launches Standalone AI Agents for Existing Contact CentersLAUNCH18 NOVMicrosoft Brings MAI Code 2 and Flash to Copilot With Promotional PricingLAUNCH

Published 19 Nov 2026