Google introduced Gemini 4 Neon on November 18, 2026, positioning it between its most demanding frontier model and its faster everyday AI offering. The company says the new model is starting to reach developers, paid Gemini subscribers and business users, bringing a one-million-token context window and published API prices of $3 per million input tokens and $15 per million output tokens.

A third route through the Gemini 4 family

Google's November 18 launch announcement describes Neon as a model for coding, scientific work and complex enterprise tasks. Google says the model is intended to balance substantial reasoning with practical task speed. That is the company's positioning, not independently proven superiority for every customer.

That distinction matters for businesses building agents. An assistant that reads an entire repository, checks documents and creates a spreadsheet may benefit from a capable model that can complete several steps consistently. But a million-token context window is a maximum input capacity, not a guarantee that every important fact remains accessible throughout the conversation or that processing the entire window is economical.

Price and availability: what is actually announced

Google advertises $3 per million input tokens, $15 per million output tokens and $0.15 per million cached input tokens, according to its launch announcement and UOL's independent reporting. Those are list prices, not measured cost per completed project. Reasoning, repeated attempts and tool execution may increase the overall bill; cached-input pricing applies only to eligible input.

Google says Neon is rolling out through its paid Gemini API and AI Studio offering, Gemini for business, Google Antigravity and the Gemini app for qualifying subscribers. UOL's November 18 account also lists these main channels. Individual subscriptions and locations should be checked rather than assuming immediate universal access.

How to interpret the benchmarks

Google highlights strong benchmark performance for Neon in its launch materials, including coding and enterprise task evaluations. UOL's coverage independently reports the announcement's headline scores. The automated editorial reviewer was unable to retrieve the model-card contents in usable form, so we do not claim its tables were independently verified by that review.

Google's reported benchmark results are not a guarantee of deployment performance. Scores depend on model settings, selected tasks and evaluation harnesses. A high score does not demonstrate that a deployed agent has appropriate permissions or can reliably finish each customer's particular job.

Google presents multi-step instruction-following and responsiveness to user corrections as priorities for Neon. These are vendor-described capabilities and require validation in real applications, including error recovery and failed tool calls.

The practical test for buyers

For engineering teams, Neon expands the choices rather than delivering a universal upgrade recommendation. Compare it against a less expensive everyday model and an established frontier model on the same real tasks. Track verified completion, latency, retries, security boundaries and total spending. A modest per-token price does not help if a workflow needs multiple failed attempts.

Nextwith.ai has separately covered Microsoft's recent model announcement and OpenAI's recent model release. They are useful market context, not direct comparable tests of Neon's real-world performance.

Sources and disclosure

Hero illustration: Original procedural artwork by Nextwith.ai editorial, 19 November 2026. Conceptual; not an official Google visualization or test result.