← In the News

Google launches Gemini 4 Neon and says its harness triples the model's SWE-sweep score

Introducing Gemini 4 Neon · Koray Kavukcuoglu, SVP, Google DeepMind · blog.google, November 18, 2026

Machine-readable Download Markdown

Neon sits between Carbon and Hydrogen in the Gemini 4 family. Google prices it at $3 per million input tokens, $0.15 per million cached input tokens and $15 per million output tokens. It is rolling out today in the Gemini API through Google AI Studio, in Gemini for business, and in Google Antigravity for paid AI Pro and Ultra subscribers.

On coding, Google reports a score of 58.6 on the Artificial Analysis Coding Agent Index, which it says ranks first overall, 65.2% on Terminal-Bench 4.1, and 46.2% on FrontierCode 2.0 at the default high setting. On SWE-sweep, Google says Neon scores 12.8% in the mini-SWE harness and "more than triples to a state-of-the-art 39.0% in Google Antigravity". Google also says Neon "cuts down on redundant file reads" and handles context compaction reliably in long sessions. At its default high thinking level, Google puts its cost per task 32% below Carbon's. There are four thinking levels.

Google's own evaluations also cover agent behaviour: shortcuts on 2.45% of tasks in its coding cheating evaluation, and 0.0% cheating on Terminal-Bench 4.0. Every benchmark, cost and behaviour figure here is Google's, or from a third-party leaderboard Google cites; we have not run the model. The launch thread had 336 points and 326 comments on Hacker News at 19:02 EST.

Why it matters: Google's own SWE-sweep figures show the harness moving one model's score threefold, so a leaderboard number says little about how Neon will do in your setup. Test it in the harness you run. Kernel's trace of Carbon, covered earlier today, found 39% of the lines Carbon read were lines it had already seen; Google's claim about fewer redundant reads is the part to check against that.