Every article on this site is written by AI from official sources and checked by a second AI. How it works

Google introduces Gemini 4 Neon at $3 per million tokens

Google on Wednesday introduced Gemini 4 Neon, a new model in its Gemini 4 family for software engineering, knowledge work and science tasks.

Google said the model is rolling out the same day in the Gemini application programming interface in Google AI Studio, in Gemini for business, and for paid Google AI Pro and Ultra subscribers in Google Antigravity and the Gemini app. The company priced Neon at $3 per million input tokens and $15 per million output tokens. Cached input tokens are $0.15 per million, which Google called a 95% discount, and the context window is one million tokens, a measure of how much material the model can take in at once.

Gemini 4 Neon price and window
ItemFigure
Input tokens$3 per million
Cached input tokens$0.15 per million
Output tokens$15 per million
Context window1 million tokens
Source: Google (The Keyword).

Google said Neon ranked third overall on the Artificial Analysis Intelligence Index with a score of 61 at its extra-high thinking level, just behind larger and slower models. The company said it ranked second on the Vals Index and sat on that index’s efficiency frontier. Google also said Neon ranked second for steerability, meaning how cleanly it adapts to corrections in the middle of a task, on a new Agent Arena Index, and posted its lowest fabrication rate yet in internal Google Antigravity testing.

For office and analytical work, Google said independent evaluations ranked Neon among the top models. The company said it placed second on the Vals Index at 70.9% and first on both a legal research bench and a tax agent bench. Google said it ranked first on most Surge AI tests covering finance, accounting, spreadsheet modeling and complex constraints, scored 73.5% on a chart-reading test called Chartography, and matched leading models at 75.8% on Agents’ Last Exam version 1.1.

Google said Neon ranked first on Zapier’s AutomationBench at 56.3% and on Andon Labs’ Vending-Bench 2, a long simulation of running a business, with a result of $16,843. The company said it debuted second globally with a score of 1,336 on Arena’s Agent Arena Index across more than 4,200 real-world agent sessions. On science and math, Google said Neon led Terminal-Bench Science at 70.5%, RiemannBench at 84.0% and HLE-Diamond at 71.3%.

For coding, Google said Neon ranked first on the Artificial Analysis Coding Agent Index with a score of 58.6, posting 65.2% on Terminal-Bench 4.1 and, when left at the usual high setting, 46.2% on FrontierCode 2.0. The company said that on a test called SWE-sweep, Neon led on one setup at 12.8% and rose to 39.0% inside Google Antigravity.

Google said Neon has four thinking levels so teams can balance intelligence against token cost. At the default high level, the company said, cost per task was 32% lower than Gemini 4 Carbon and about 50% to 80% lower than top competing models on the Vals index and Agent Arena. At the extra-high level, Google said, cost per task was still at least 36% below competing models at their highest setting.

Google also described a cyber version meant to write more secure code and to find, check and patch flaws. The company said it beat Carbon across easy, medium and hard challenges on AutoBaxBench and reached 76% on the hard tier, scored 51.26% on WebExploitBench version 1.2 at half the cost of competing large models, and reached 90.5% on SEC-bench Pro at about half the cost per challenge. Google said it used half the tokens of Carbon on ExploitGym and reached 70.7% on Collinear’s CWE-bench version 2, a 3.4% uplift over Carbon. That version will go to trusted defenders in the coming days through the Fairwind Program, and eligible members can ask a Cloud account team for access, Google said.

On safety, Google said Neon keeps Carbon-level protections against severe misuse and prompt injection, with real-time safeguards against cyber and biosecurity misuse, while turning away less ordinary work. The company said it refused 1.98% of tasks on Terminal-Bench 4.0 and none on Terminal-Bench-Science. Google said internal monitors stayed below Carbon’s already low alert baselines, with no cheating on Terminal-Bench 4.0, shortcuts on 2.45% of tasks in a coding test, a 0.09% misaligned-action rate and a 99.2% pass rate on an honesty evaluation. It said under-service fell 43% to 56% and overreach fell 15% to 35% compared with Carbon, depending on effort. On a Gray Swan test of hidden malicious instructions, Google said Neon with safeguards at its highest thinking level resisted every attack in that benchmark.

In the Gemini app, Google said, Pro and Ultra subscribers can select Neon for writing, reading uploads such as spreadsheets and PDFs, research that draws on web sources, and faster multi-step jobs. Developers can use the paid Gemini API in Google AI Studio and Android Studio, and paid Pro or Ultra subscribers can try it in Google Antigravity and AI Studio, the company said. The post was by Koray Kavukcuoglu, senior vice president of Google DeepMind and chief AI architect at Google. The page said the article was generated by Google AI and that generative AI is experimental.

  1. Introducing Gemini 4 Neon — Google (The Keyword)
How this article was made
Written with
Grok 4.7 (xAI)
Fact-checked by
Claude Sonnet 5.5 (Anthropic) — passed after revision
Generated
Nov. 18, 2026, 4:42 p.m. ET
Sources used
1 official source, listed above
Article type
Standard

The AI wrote only from the sources listed. No human reporter was involved. Read our standards or compare the models in the AI Report Card.