Google introduced Gemini 4 Neon on November 18, 2026, as a model for complex programming, knowledge work and scientific tasks. The company says it can edit code across multiple files and process different types of documents. Its rollout has begun across developer tools, business products and the Gemini app.
Code edits and subagents
Google says Neon can debug and refactor multiple files across repositories using targeted edits. In Google Antigravity, it can work independently or direct Hydrogen subagents running in parallel.
For long-running tasks, the company says the model reduces repeated file reads, shares progress as it works and compacts its context. Neon offers four reasoning levels; according to Google, the xhigh setting adds deeper multi-step reasoning for the most demanding engineering, analytical and scientific tasks.
Documents and multi-step tasks in Gemini
Google says the model can combine spreadsheets, PDFs, charts, audio and video within a single one-million-token context window and turn them into structured outputs. In the Gemini app, it is designed to read uploaded photos more accurately, prepare reports and summaries, and use related information from Gmail and Google Docs.
Tasks in Gemini can now use Neon for multi-step jobs. Google gives examples including analyzing bills and comparing receipts with other documents when requesting a refund.
Coding and analytical test results
Google says Neon scored 61 on the Artificial Analysis Intelligence Index at the xhigh setting, placing third overall. It ranked second on the Agent Arena Index for its ability to adapt to user corrections during a task.
On the SWE-sweep coding test, Google reports that the same model scored 12.8% with the mini-SWE test harness and 39.0% in the Google Antigravity environment. The company also reports 65.2% on Terminal-Bench 4.1 and 46.2% on FrontierCode 2.0 at the default high reasoning level.
For analytical work, Google reports a score of 73.5% on Chartography, which tests the reading of complex professional charts, and second place on the Vals Index with 70.9%. Its reported scientific and mathematical results include 70.5% on Terminal-Bench Science and 84.0% on RiemannBench.
A cybersecurity variant
Google also introduced Gemini 4 Neon for cyber. The company says it produces more secure code than Carbon on all three AutoBaxBench difficulty levels, scoring 76% on the hardest. On CWE-bench v2, which tests vulnerability fixes, it scored 70.7%—an improvement of 3.4% over Carbon, according to Google.
Eligible Fairwind Program members are due to begin receiving access to the security variant over the following days. People interested in the program should contact their Google Cloud team.
Where Gemini 4 Neon is available
Developers are gaining access through the paid Gemini API tier in Google AI Studio and Android Studio. The model is rolling out to businesses through Gemini for business and Google Antigravity. In the Gemini app, it is available to Google AI Pro and Ultra subscribers, who can also try it in Antigravity and AI Studio.
The API costs $3 per million input tokens, $0.15 per million cached tokens and $15 per million output tokens; Google also says Neon costs 32% less per task than Carbon on the Artificial Analysis index at the default high setting.



