Google Launches Gemini 4 Neon as Its New Default AI Model

Google signage at a Google office campus, the company behind the newly launched Gemini 4 Neon AI model
Photo: Wikimedia Commons, CC BY-SA 4.0.

Last updated: November 19, 2026, 08:00 IST.

Google began rolling out Gemini 4 Neon on November 18, a new artificial intelligence model built for coding, enterprise work and scientific research. The company says Neon matches or beats its top tier Carbon model on several benchmarks while costing up to 36 percent less per task, and it now replaces Argon as the default model in the Gemini app.

Neon sits between Google’s faster Hydrogen model, meant for everyday tasks, and Carbon, reserved for the hardest, longest running projects, according to Google’s own announcement. It reads a 1 million token context window, can work with spreadsheets, PDFs, charts, audio and video, and can act inside Google’s Antigravity coding tool as an orchestrator that directs multiple Hydrogen subagents at once.

The release follows Carbon and Hydrogen, which Google shipped the week before, 9to5Google reported. The outlet said Neon now replaces Argon entirely in the Gemini app’s model picker, and that Gemini 3.1 Pro is being retired at the same time.

How Neon’s benchmarks compare

Google’s own results put Neon third overall on the Artificial Analysis Intelligence Index, scoring 61 at its highest “xhigh” thinking setting, just behind larger, slower models. On the Vals Index, a benchmark built around real world professional tasks, Neon ranked second with a score of 70.9 percent, and placed first on both the Legal Research Bench and the Tax Agent Bench, according to Google’s announcement.

For coding specifically, Google says Neon tops its own Artificial Analysis Coding Agent Index with a score of 58.6, and reached 39.0 percent on the SWE-sweep benchmark when run inside Antigravity, a result the company calls state of the art for that harness. VentureBeat reported that Neon’s coding performance is comparable to Anthropic’s Opus 5.5 while its API tokens cost 25 percent less once Google’s current promotional pricing ends.

On science and reasoning tasks, Google reported Neon leads its model family on Terminal-Bench Science (70.5 percent), RiemannBench (84.0 percent) and HLE-Diamond (71.3 percent). For enterprise and agent workflows, the company cited a first place score on Zapier’s AutomationBench (56.3 percent) and on Andon Labs’ Vending-Bench 2, where Neon’s simulated business ended the run with 16,843 dollars.

What it costs and who can use it

Neon is priced at 3 dollars per million input tokens and 15 dollars per million output tokens through the Gemini API, with cached input tokens discounted 95 percent to 0.15 dollars per million, per Google’s pricing page. Argon and Carbon currently cost 2 dollars per million input tokens and 10 dollars per million output tokens under a promotion that ends November 30, after which both rise to 4 dollars and 20 dollars respectively, 9to5Google reported.

ModelInput $ / 1M tokensOutput $ / 1M tokensRole in the lineup
Neon3.0015.00New default, mid tier
Argon / Carbon (through Nov 30)2.0010.00Promotional pricing
Argon / Carbon (from Dec 1)4.0020.00Standard pricing

Once the promotion ends, Neon will cost 25 percent less than Argon or Carbon for the same token volume, Google’s own pricing figures show. The model is available now in Google AI Studio and Android Studio, through the Gemini API, inside Google Antigravity for paid AI Pro and AI Ultra subscribers, and in the Gemini app’s model dropdown for Pro and Ultra subscribers.

Google is also sending a separate version, called Neon for cyber, to eligible members of its Fairwind Program over the next few days, aimed at security teams rather than general developers.

Safety numbers Google reported

Alongside the performance figures, Google published safety metrics for Neon. The company says the model has a 2.45 percent “shortcut” rate on its internal coding cheating evaluation, which it calls the lowest of any Gemini 4 model released so far, and a 0.09 percent rate of what it labels misaligned actions.

Neon also matched Carbon’s 0.0 percent cheating rate on Terminal-Bench 4.0, passed 99.2 percent of an agent honesty evaluation, and recorded a 0.0 percent attack success rate on Gray Swan’s indirect prompt injection benchmark at its highest thinking setting, according to Google’s announcement.

A separate version built for cyber defense

On cybersecurity benchmarks specific to Neon for cyber, Google reported a 76 percent score on the hard tier of AutoBaxBench, 51.26 percent on WebExploitBench, which it calls state of the art for that test, and 90.5 percent on SEC-bench Pro. On CWE-bench v2, Google said Neon improved 3.4 percentage points over Carbon, to 70.7 percent.

FAQ

What is Gemini 4 Neon? It is Google’s newest AI model, released November 18, positioned between its Hydrogen model for quick everyday tasks and its Carbon model for the hardest, longest projects. Google built it for coding, enterprise workflows and scientific research, and it now replaces Argon in the Gemini app.

How much does Neon cost to use through the API? 3 dollars per million input tokens and 15 dollars per million output tokens, with a 95 percent discount on cached input tokens, per Google’s own pricing page.

Where can I use it? Through the Gemini API in Google AI Studio and Android Studio, inside Google Antigravity for paid AI Pro and Ultra subscribers, and in the Gemini app’s model dropdown for Pro and Ultra subscribers. A separate cyber defense version is going out to eligible Fairwind Program members.

What to watch next: whether Neon’s self reported benchmark scores hold up once independent trackers like Artificial Analysis publish their own measurements.

Source: Google (Gemini 4 Neon announcement), 9to5Google, VentureBeat.