aipopdb

22 Nov 2026  ·  10,137 works · 1,192 news indexed

News / Google releases Gemini 4 Neon, a $3/$15 mid-tier model for the Gemini app and API

Google releases Gemini 4 Neon, a $3/$15 mid-tier model for the Gemini app and API

AIPopDB Newsroom · via blog.google · · 5 min read · Comment

Read the original ↗
Image for Google releases Gemini 4 Neon, a $3/$15 mid-tier model for the Gemini app and API

Image: Google (The Keyword)

Watch on YouTube ↗

Google released Gemini 4 Neon on 18 November 2026, a mid-priced model that sits between Gemini 4 Hydrogen and Gemini 4 Carbon in the Gemini 4 family. Developers can reach it through the Gemini API in AI Studio, businesses through Gemini for business, and paid Pro and Ultra users through the Antigravity coding environment. Pro and Ultra subscribers also get it in the Gemini app, where it takes the place of Gemini 4 Argon while Gemini 3.1 Pro is being retired.

The app's model picker is reorganised around the new line-up. Hydrogen is labelled for everyday help, Neon for deeper analysis and Carbon for code and complex work, and the older 3.5 Flash-Lite becomes an Auto option for paying users. With three tiers now labelled by job, a subscriber no longer has to guess which model suits which request. Neon is Google's fourth new Gemini 4 model in four weeks, arriving after Carbon and Hydrogen a week earlier, and Argon's own app rollout came on 13 October.

On price, Neon costs $3 per million input tokens and $15 per million output tokens, with cached input at $0.15. That is twice Hydrogen's rate, which The Decoder gives as $1.50 and $7.50. Carbon and Argon sit at $2 and $10 during an introductory period, after which they rise to $4 and $20, which would make Neon about a quarter cheaper than either. The outlets disagree on when that step-up lands: 9to5Google says after 30 November, The Decoder says 1 December. It has a one million token context window and four thinking levels, from Low to Extra High, takes text, images, video, audio and PDFs, and supports computer use, code execution, file search and grounding with Search and Maps. It does not generate images or audio, so it is a reasoning and agent model rather than a creative one.

“Neon also excels at multi-step instruction following and builds on our ongoing progress to curb fabrication, recording our lowest fabrication rate yet in internal Google Antigravity testing.”

Google, in its announcement post, in its announcement post · blog.google ↗

The model comes from Google DeepMind, and the announcement post is the fullest account of what it can do. Google frames Neon as the point where frontier-level reasoning meets quick end-to-end task completion, aimed at software engineering, knowledge work and science. 9to5Google reads it more bluntly: in its assessment Neon matches or beats Carbon on coding, reasoning and agentic tasks, at the speed of Google's Flash-class models and at a lower cost. Antigravity, Google's coding environment, is where the company shows this off most. The pitch, in other words, is that people who reached for the top-tier model for hard jobs can now get similar results from a cheaper, faster one.

Most of Google's case rests on benchmarks it cites itself. Neon scores 61 on the Artificial Analysis Intelligence Index at its highest thinking level, ranking third, and second on the Vals Index at 70.9%. It is first on the Legal Research Bench and Tax Agent Bench, first on Zapier's AutomationBench at 56.3% and first on Andon Labs' Vending-Bench 2, where it ended with $16,843. It is second on Arena's new Agent Arena Index with 1336 across more than 4,200 agent sessions. Google also lists 70.5% on Terminal-Bench Science, 84.0% on RiemannBench and 71.3% on HLE-Diamond.

Further results cover agent work and charts: 73.5% on Chartography and 75.8% on version 1.1 of Agents' Last Exam. Because these are Google's own figures, taken from its announcement and repeated by 9to5Google, they describe how the company chose to present the model rather than an independent verdict, although some of them come from third-party indexes and leaderboards.

9to5Google's cover graphic for its Gemini 4 Neon report

9to5Google's cover graphic for its Gemini 4 Neon report

Image: 9to5Google

Coding is a particular focus. Google says Neon leads the Artificial Analysis Coding Agent Index with 58.6, and gives 65.2% on Terminal-Bench 4.1 and 46.2% on FrontierCode 2.0. On SWE-sweep it reports 12.8% with the mini-SWE harness and 39.0% inside Antigravity, so the setup a model runs in changes the result a good deal. At the default high thinking setting, Google puts its cost per task 32% below Carbon's on the Artificial Analysis index. Its demos in Antigravity include a WebGL2 ski simulator on Mont Blanc terrain, a 3D recreation of Turing's 1936 Cambridge study and a 100,000-cell spreadsheet built with over 500,000 formulas, each meant to show Neon working through a long, multi-part job without being walked through it step by step.

Google also published safety figures, all from its own evaluations of how the model behaves when it works through tasks on its own. In its tests Neon declined 1.98% of Terminal-Bench 4.0 tasks, showed no cheating there, scored 99.2% on an agent honesty evaluation and had a 0.09% misaligned-action rate. A separate cyber-focused version, Gemini 4 Neon for cyber, is due for trusted defenders in the coming days through the Fairwind Program, which Google says will be handled separately from the general release, with 76% on the hard tier of AutoBaxBench and 90.5% on SEC-bench Pro.

The release continues a fast run of launches. Argon was announced in late September and reached the Gemini app for Pro and Ultra subscribers on 13 October, Carbon and Hydrogen arrived the week before Neon. Looking ahead, the introductory pricing on Carbon and Argon is close to its end, and the cyber version is meant to reach defenders within days. Nothing in the coverage gives a date for anything beyond that.

The Decoder adds the comparison to rivals: if Carbon and Argon are Google's answers to OpenAI's Astra or Anthropic's Fable, it suggests Neon is closer to an Opus or Sol equivalent. It also reports an output ceiling of up to one million tokens, a figure Google's post does not give; the post names only the million-token context.

Comments