Change language to
0:00

Google’s new Gemini 4 Neon targets complex coding and knowledge work at $3 per million input tokens and $15 per million output tokens.

Through the API, cached inputs cost $0.15 per million tokens, a 95% discount on the standard input rate. Neon can handle up to 1 million tokens of context.

Koray Kavukcuoglu, SVP of Google DeepMind and Google’s chief AI architect, announced Neon as a lower-cost option for completing complex tasks end to end. It sits between Carbon, built for the hardest long-running projects, and Hydrogen, built for fast everyday work.

Subscribe to our Newsletters for more Tech Stories

Who can use Gemini 4 Neon?

The rollout began on 18 November in the Gemini API in Google AI Studio and in Gemini for business, where users can access it through the API or model dropdown. Google AI Pro and Ultra subscribers can pick Neon in the Gemini app, and paid Google AI Pro and Ultra subscribers can use it in Google Antigravity.

Google hasn’t specified UAE or wider GCC availability or local pricing. In the run-up, Google DeepMind’s chief said Gemini 4 was nearly ready. Hydrogen has reached Google Search for all users, not just Ultra subscribers.

Google DeepMind has also published a model card for Neon.

Google’s benchmark claims for Neon

Google’s own figures put Neon first overall on the third-party Artificial Analysis Coding Agent Index, with a score of 58.6. Neon scored 65.2% on Terminal-Bench 4.1 and 46.2% on FrontierCode 2.0, the latter at its default high thinking level.

Google’s figures put Neon third overall on the Artificial Analysis Intelligence Index at its xhigh setting, with a score of 61 behind larger, slower models. It placed second on the Vals Index at 70.9%. It also placed second on the Agent Arena Index for steerability, which measures how well it adapts to mid-task corrections.

Google says Neon costs 32% less per task than Carbon at its default high setting.

Neon shows its highest intelligence score here, while still costing 36%+ less per task than competing models at their highest setting.

Koray Kavukcuoglu, SVP of Google DeepMind and chief AI architect, on Neon’s xhigh setting

Neon matched Carbon’s 0.0% cheating rate on Terminal-Bench 4.0 and refused 1.98% of tasks on that test. It took shortcuts on 2.45% of tasks in the separate in-house coding evaluation, the lowest rate of any Gemini 4 model.

A separate rollout for cyber defenders

Google says the cyber version of Neon produced more secure code than Carbon on AutoBaxBench and scored 70.7% on CWE-bench v2. Eligible defenders will start getting access over the next few days through Fairwind, the same programme Google used to roll out Gemini 4 Argon to defenders first.