Google released Gemini 4 Neon on November 18, a model it places between its Carbon and Hydrogen tiers, charging $3 for each million input tokens and $15 for each million output tokens. Google says Neon scores 61 on the Artificial Analysis Intelligence Index at its xhigh thinking setting, third overall and just behind larger, slower models. The pitch is speed and price as much as raw ability.

Google’s own description of the lineup explains the slot. Carbon takes the hardest, long-horizon projects, Hydrogen handles fast everyday execution, and Neon is the middle tier for complex, high-frequency work. Koray Kavukcuoglu wrote the announcement on the Google blog.

Cached input costs $0.15 per million tokens, which Google calls a 95 percent discount, and the context window holds 1 million tokens. Neon is rolling out in the Gemini API in Google AI Studio and in Gemini for business. Paid Google AI Pro and Ultra subscribers get it in Google Antigravity, and Pro and Ultra subscribers can pick it in the Gemini app.

The coding numbers are where Google leans hardest. It reports first place on the Artificial Analysis Coding Agent Index at 58.6, 65.2 percent on Terminal-Bench 4.1, and 46.2 percent on FrontierCode 2.0 at the default high setting. On Vals Index, Neon ranks second at 70.9 percent. On Arena’s new Agent Arena Index it debuts second with 1336, over a sample of more than 4,200 live agent runs.

One result says more about the harness than the model. On SWE-sweep, Neon scores 12.8 percent on the mini-SWE harness and 39.0 percent inside Google Antigravity, more than triple. Google reads that as proof of what Antigravity unlocks. A buyer can read it another way: the score you get depends on the tooling around the model, and the best figure comes from Google’s own.

The cost claims carry the most weight for operators. Google says Neon at high thinking costs 32 percent less per task than Carbon, and roughly 50 to 80 percent less than top competing frontier models. At xhigh, it says the saving is still 36 percent or more against competitors at their highest settings. The announcement does not name those competing models, so the comparison cannot be reproduced from the post alone.

Outside coding, Google reports leading results on Legal Research Bench and Tax Agent Bench, 56.3 percent on Zapier’s AutomationBench, and $16,843 on Andon Labs’ Vending-Bench 2, a simulated business run. On science and math it claims the top spot on Terminal-Bench Science (70.5 percent), RiemannBench (84.0 percent), and HLE-Diamond (71.3 percent).

A cyber variant, Gemini 4 Neon for cyber, starts reaching vetted security teams within days via the Fairwind Program. Google reports 90.5 percent on SEC-bench Pro for finding vulnerabilities and 51.26 percent on WebExploitBench v1.2 while spending about half what rival large models do. Eligible members request access through their Cloud account team.

The safety figures are Google’s internal evaluations. It reports a 0.09 percent misaligned action rate on its agentic safety suite, a 99.2 percent pass rate on its agent honesty test, and refusals on 1.98 percent of Terminal-Bench 4.0 tasks. With real-time safeguards at the highest thinking level, Google says Neon resisted every attack in Gray Swan’s prompt injection benchmark. It also calls the fabrication rate in internal Antigravity testing its lowest yet. The post links a model card but offers no independent replication.

Third-party indices anchor the headline ranking, which is more than most launch posts offer. The cost-per-task gap against Carbon is the figure a team can verify cheaply: replay a week of real coding or analysis jobs at high thinking and compare the bills before deciding which tier becomes the default.

Google, “Introducing Gemini 4 Neon,” blog.google, November 18, 2026.