top of page

TECH NEWS · AI

Google Gemini 4 Neon: frontier smarts, faster finish — at $3 per million tokens

Google DeepMind has launched Gemini 4 Neon, the fourth model in the Gemini 4 family — a fast frontier model built for complex software engineering, knowledge work and science. It costs $3 per million input tokens, carries a 1-million-token context window, and scores 61 on the Artificial Analysis Intelligence Index, third overall. This is a news article, not a review: no one at ArcTech Asia has tested Neon hands-on, and there is no verdict or score here. What follows is what Google announced, what it costs, and where you can actually use it.

By Sky Oh, ArcTech Asia

Input price — $3 per million tokens

Output price — $15 per million tokens

Context — 1 million tokens

Benchmark — 61, #3 on the Artificial Analysis index

01 — What Neon is

Google DeepMind announced Gemini 4 Neon on Tuesday 18 November 2026. It is the fourth model in the Gemini 4 family, after Argon and Carbon (10 November) and Hydrogen (12 November) — and the one Google is pitching on speed as much as smarts: frontier-level intelligence that finishes long, multi-step tasks quickly, rather than the biggest, slowest model in the range. The target is the work that now dominates AI usage: complex software engineering, knowledge work and science workflows, where a model has to plan, run tools and iterate, not just answer a single question.

Bringing frontier-level intelligence with fast end-to-end task completion across complex software engineering, knowledge work, and science workflows.
— Google DeepMind, on Gemini 4 Neon

02 — What it is for

Neon is aimed at the agentic end of the spectrum: coding agents that work through a task in many steps, data analysis across large documents and datasets, and scientific research where the model has to reason through a problem end to end. The pitch is the balance. The largest models still lead on raw scores, but they are slower and more expensive to run. Neon is Google's attempt to hold most of that capability while finishing the job faster — the tier you reach for when the task is long and the clock matters.

03 — The numbers

On the Artificial Analysis Intelligence Index, Neon scores 61 at xhigh thinking — third overall, behind larger, slower models. That is the headline number, and it is Google's own reporting of the benchmark. Two other figures are reported, not from Google's own announcement, so treat them as reported: Neon is said to rank first on the Coding Agent Index (58.6) and second on the Vals Index (70.9%). Both point the same way — Neon is strongest where the work is agentic and code-heavy — but neither figure comes from the official blog, and this article does not present them as Google's numbers.

04 — Price and context

Neon is priced at $3 per million input tokens, $0.15 per million cached input tokens, and $15 per million output tokens, with a 1-million-token context window. There is no Singapore-specific price — the API pricing is in US dollars, the same worldwide. The pricing puts Neon in the fast-frontier tier: cheaper to run per task than the biggest models, with the context window to hold a whole codebase or a long research paper in one go. Third-party estimates of how much cheaper per task than Carbon have disagreed with each other, and neither is official, so this article makes no claim on that comparison.

05 — Where to use it

Neon started rolling out on 18 November in the Gemini API and Google AI Studio, in Gemini for business, in Google Antigravity for paid Pro and Ultra subscribers, and in the Gemini app for Pro and Ultra subscribers. What Google has not said is whether Neon comes to free users, or to AI Mode in Search. Until it does, do not assume either: for now, Neon is a paid-tier model.

06 — The same day: Genie 4

Neon was not the only DeepMind launch on 18 November. The same day brought an update to Project Genie, with Genie 4 — the tool that generates interactive 3D worlds from text prompts — now available for eligible Ultra subscribers, while Genie 3 remains available. Keep the two distinct. Genie is the interactive-worlds tool; Neon is the model. They arrived together, but they are not the same product.

07 — The catch

Three things to know. First, the paywall: Neon is a paid-tier model for now, with no word from Google on free users or AI Mode in Search. Second, the headline scores are Google's own reporting and the third-party figures labelled as reported — not independent lab tests. Third, the real test is not a benchmark: it is whether Neon holds up in actual coding-agent use, where the work is long, the context is large and the model has to keep going without losing the thread. What to watch: whether Neon reaches free users and AI Mode, and how it compares in real use against the bigger models it sits just behind on the index.

bottom of page