Google launched Gemini 4 Neon on Wednesday, a new mid-tier AI model built for coding, research and office work that the company says finishes complex jobs faster and cheaper than its flagship model.
Neon is the fourth model in Google’s Gemini 4 family to arrive in a matter of weeks, and it lands in the middle of a crowded stretch for the AI industry. OpenAI released two new models days earlier, and Anthropic priced the largest stock market debut on record on the same day. Google’s answer is a model aimed at the work most customers actually run every day, rather than the hardest problems its biggest model is built for.
Key Points
- Google launched Gemini 4 Neon on 18 November, a mid-tier model sitting between Gemini 4 Hydrogen and Gemini 4 Carbon.
- It costs $3 per million input tokens and $15 per million output tokens, with a one-million-token context window.
- Google reports a score of 61 on the Artificial Analysis Intelligence Index, third overall, and first place on the coding agent index.
- The model is available now through the paid Gemini API and to Google AI Pro and Ultra subscribers. Free users do not get it.
- Until 30 November, Carbon’s promotional price makes Neon the dearer option per token. From 1 December, Neon becomes 25% cheaper.
What happened
Google DeepMind announced the model on Wednesday in a blog post by Koray Kavukcuoglu, its senior vice president and chief AI architect. Neon is built for long, multi-step work: coding, analysis and science, where the model has to follow evidence through several stages and hand back something a person can use.
The model sits between the two Gemini 4 models Google released earlier this month. Carbon handles the hardest long-running projects. Hydrogen is the fast, cheap option for everyday tasks. Neon is meant to balance depth and speed, for work that needs real reasoning but runs often enough that cost and pace matter.
It carries a one-million-token context window, so it can take in spreadsheets, PDFs, charts, audio and video together and turn them into structured results. Four thinking levels let users trade intelligence against token spend, with the highest level, called xhigh, reserved for the hardest jobs.
Neon costs $3 per million input tokens and $15 per million output tokens on the paid tier of the Gemini API. Cached input costs $0.15 per million tokens, a 95% discount. It is rolling out in Google AI Studio and Android Studio, in Gemini for business, in Google’s coding tool Antigravity for paid Pro and Ultra subscribers, and in the Gemini app for Pro and Ultra subscribers.
Free users do not get Neon, and neither do subscribers on the cheaper AI Plus plan. In the Gemini app, Neon takes the place of Argon in the model picker, and the older Gemini 3.1 Pro is being retired, according to 9to5Google. Carbon stays limited to Ultra subscribers. GitHub has also added Neon to Copilot for its Pro+, Max, Business and Enterprise plans.
A security-focused version, Gemini 4 Neon for cyber, will reach trusted defenders in the coming days through Google’s Fairwind Program. Google says it writes more secure code than Carbon and finds flaws faster, with state-of-the-art results on several security tests it cites.
Why it matters
The headline numbers come from Google’s own tests, and they need that label. On the Artificial Analysis Intelligence Index, Neon scores 61 at its highest thinking level, third overall behind OpenAI’s GPT-6.2 Astra and Anthropic’s Claude Fable 5.2. Artificial Analysis confirmed the third-place ranking. On the Vals Index, which grades finance, legal and tax work, Neon scores 70.9% and ranks second. It tops the Artificial Analysis Coding Agent Index with 58.6.
In coding tests, the gap to the biggest models is small. On Cognition’s FrontierCode 2.0, Google reports 46.2% for Neon against 47.3% for Anthropic’s Claude Opus 5.5. On Zapier’s AutomationBench, Neon reaches 56.3%, ahead of Opus at 42.5% and GPT-6.2 Sol at 41.3%. On Terminal-Bench Science, a test of scientific reasoning, Google reports 70.5%.
Speed is part of the pitch. Benchmark tracker Priors measured Neon streaming about 110 tokens a second at its highest setting, against 59 for Carbon, or roughly 1.9 times faster. Google says a customer, Glean, found Neon matched or beat Carbon on quality in early tests. Thai Tran, Glean’s AI product lead, said in Google’s announcement: “In our early evaluations, it matched or exceeded Gemini 4 Carbon on overall quality at 3x the speed and 60% of the cost.”
The price story has a twist that Google’s announcement does not spell out. Carbon’s $2 input and $10 output rates are promotional until 30 November. Until then, Neon costs more per token than the flagship it is meant to undercut. From 1 December, Carbon rises to $4 and $20, and Neon becomes 25% cheaper per token without moving at all.
Cost per task tells a similar story with a smaller gap than Google claims. Google says Neon costs 32% less per task than Carbon at its default high setting. Artificial Analysis, measuring at the highest setting where Neon earned its 61, put the advantage at about 17%. The two figures use different settings, so buyers who run Neon at full effort will save less than the headline suggests.
On safety, Google says Neon resisted every attack in Gray Swan’s indirect prompt injection test at its highest thinking level, where harmful instructions are planted in pages, emails or documents the model reads. The company also says Neon has its lowest fabrication rate yet in internal Antigravity testing. The full safety and security report for Neon has not been published yet. The model card marks it as coming soon.
What happens next
The date that matters is 1 December, when Carbon’s promotional pricing ends and Neon’s price advantage becomes real. Until then, developers paying per token will find Carbon the cheaper choice, and the decision turns on speed and availability rather than price.
The cyber version reaches trusted defenders through the Fairwind Program over the coming days, and Google’s full safety report is still to come. Independent testing will also matter: the benchmark figures so far are Google’s own, and outside labs have yet to run Neon through their own checks.
The launch caps a frantic week for the industry. OpenAI released GPT-6.2 Astra and Sol days earlier, and Anthropic priced its record $79.4 billion listing on the same day Neon arrived. With four Gemini 4 models now in the family, Google’s next test is whether customers pick the middle model, or keep reaching for the biggest one.
Sources: Google, VentureBeat, The Verge, 9to5Google, Artificial Analysis. Time: 19 November 2026.



