Google releases Gemini 4 Neon AI model

Google has announced the release of a new artificial intelligence model, Gemini 4 Neon. The model is designed for software engineering, analytics and scientific research and is positioned as a high-performance solution for complex multi-step tasks.

Performance and accuracy rankings

Gemini 4 Neon ranks highly on several respected benchmarks: third place in the Artificial Analysis Intelligence Index with a score of 61 at the “xhigh thinking” level, and second place on Val’s Index, where the model sits on the Pareto frontier. In the new Agent Arena Index Neon took second place for steerability, and Google’s internal Antigravity tests showed the lowest hallucination rate among the models tested.

server room with glowing AI charts and code

Pricing, context window and availability

The model is available at a rate of $3 per million input tokens (15 cents per million cached input tokens, a 95% discount) and $15 per million output tokens. Gemini 4 Neon has a context window of 1 million tokens and is already available via the Gemini API through Google AI Studio, and also in Gemini for Business via the API and model selection.

Subscribers to Google AI Pro and Ultra will get access to the model in Google Antigravity, and Gemini Pro and Ultra users will be able to use it in the Gemini app.

Industry applications and speciality benchmarks

Gemini 4 Neon can process data from multiple sources - spreadsheets, PDFs, audio and video - within a single context window. On Val’s Index it scored 70.9%, took first place in the Legal Research Bench and Tax Agent Bench, and also leads in the Surge AI categories related to finance, accounting and spreadsheet modelling.

visualization of neural network processing data

Code performance and cybersecurity

For coding tasks, Neon ranks first in the Artificial Analysis Coding Agent Index (58.6), scores 65.2% on Terminal-Bench 4.1 and 46.2% on FrontierCode 2.0, and works effectively with multi-file debugging and refactoring. In Google Antigravity using /swarm the model created a photorealistic 3D laboratory, a 3D ski simulator, an interactive 3D reconstruction of Alan Turing’s study and a complex spreadsheet with advanced calculations.

In cybersecurity Neon generates safer code than Carbon across all difficulty categories on AutoBaxBench (76%). The model achieves the best results on WebExploitBench v1.2 (51.26%) for vulnerability discovery, reduces cost compared with other models by roughly 50% and reaches 90.5% on SEC-bench Pro while costing roughly half as much. On ExploitGym Neon uses half the tokens of Carbon, and on CWE-bench v2 it reaches 70.7%, improving on Carbon by 3.4%.

Iryna Pavlova

Iryna Pavlova

Deputy Editor-in-Chief

Deputy Editor-in-Chief of Concord Fintech. She runs the news feed: choosing topics, fact-checking and preparing stories on the economy, business and government decisions that affect Ukrainians' money. She makes sure every story has a verified primary source.

All author's materials (1841)

Leave a reply