Google Launches Gemini 4 Neon: Frontier Brains, Flash Speeds, $3 Pricing
Key Takeaways
- Google has launched Gemini 4 Neon, announced on 18 November 2026 by Koray Kavukcuoglu, Senior Vice President at Google DeepMind and Chief AI Architect at Google, in a post on Google's official blog titled "Introducing Gemini 4 Neon".
- Neon is priced at $3 per million input tokens ($0.15 per million cached input tokens) and $15 per million output tokens, with a 1 million-token context window and four configurable thinking levels.
- Google says Neon ranks #3 overall on the Artificial Analysis Intelligence Index with a score of 61 at its highest thinking level, and #1 on the Artificial Analysis Coding Agent Index (58.6).
- The model is rolling out now in the Gemini API in Google AI Studio, Gemini for business, Google Antigravity for paid Pro and Ultra subscribers, and the Gemini app for Pro and Ultra subscribers.
- Neon sits between Carbon (Google's hardest, long-horizon model) and Hydrogen (its fast everyday model). Reports of a "Gemini 4 Krypton" model are unconfirmed rumours — Google has announced no such model.
What Google actually announced
The announcement is short on hype and long on benchmarks, which is itself a signal. Google's blog post, published 18 November 2026, introduces Neon as a model "bringing frontier-level intelligence with fast end-to-end task completion across complex software engineering, knowledge work, and science workflows". That sentence is doing a lot of work: it positions Neon not as the smartest model Google has ever built, but as the smartest model it can run fast and cheaply enough to use all day.
This is the fourth Gemini 4 model in seven weeks, according to Artificial Analysis, and the cadence matters. Google is no longer launching one flagship a year and iterating around it. It is shipping a family of models, each tuned for a different point on the trade-off between intelligence, speed and cost, and it is shipping them weeks apart. Neon is the latest member of that family, and its job is the middle of the range: harder work than Hydrogen can handle, faster and cheaper than Carbon.
The rollout began on the day of the announcement. Developers can build with the Gemini API paid tier in Google AI Studio and Android Studio; enterprises can use it in Gemini for business via the API or the model drop-down; paid Google AI Pro and Ultra subscribers get it in Google Antigravity and in the Gemini app. A separate variant, Gemini 4 Neon for cyber, is rolling out to trusted defenders in the coming days as part of Google's Fairwind Program.
Where Neon sits in the Gemini 4 family
Google's own framing is the clearest guide to Neon's place. In the announcement, the company says Carbon tackles "your hardest, long-horizon projects" and Hydrogen "handles fast everyday execution", while Neon "balances depth and efficiency" and "delivers frontier-level reasoning for complex, high-frequency workflows faster and more cost efficiently".
The Antigravity team put it more plainly in its own blog post: Neon "hits the sweet spot between Gemini 4 Carbon for your deepest long-horizon projects and Gemini 4 Hydrogen for rapid everyday execution".
That positioning is worth understanding because it explains the pricing. Neon is not Google's cheapest model and not its most capable. It is the model Google wants developers to reach for by default — the one that is smart enough for serious work and cheap enough to run constantly. If that sounds like a deliberate product strategy rather than a research milestone, that is because it is.
For background on the family, see our explainers on Gemini 4 Carbon and Gemini 4 Hydrogen pricing and benchmarks.
The benchmarks Google chose to highlight
Google's announcement is built around third-party benchmarks, and the numbers are specific enough to be checked. The headline figure is a score of 61 on the Artificial Analysis Intelligence Index at Neon's highest ("xhigh") thinking level, ranking #3 overall — behind larger, slower models, by Google's own admission. On the Vals Index, Neon ranks #2 with a score of 70.9 percent, and it takes #1 on both the Legal Research Bench and the Tax Agent Bench. On the Arena Agent Arena Index, it debuts at #2 globally with a score of 1336 across more than 4,200 real-world agent sessions, and ranks #2 on steerability — how cleanly the model adapts to mid-task user corrections.
For coding, the numbers are stronger. Neon ranks #1 overall on the Artificial Analysis Coding Agent Index with a score of 58.6, scores 65.2 percent on Terminal-Bench 4.1, and 46.2 percent on FrontierCode 2.0 at its default high setting. On SWE-sweep, it leads on the mini-SWE harness at 12.8 percent and more than triples to 39.0 percent in Google Antigravity, which Google presents as evidence of what the Antigravity harness unlocks. For production code, Google says Neon handles multi-file debugging and refactoring across repositories with "targeted, surgical edits", and that it is engineered for long-running agent workflows — cutting down redundant file reads, sharing progress as it works, and handling context compaction reliably in long sessions.
For science and mathematics, Google reports best-in-class results on Terminal-Bench Science (70.5 percent), RiemannBench (84.0 percent) and HLE-Diamond (71.3 percent). On enterprise and analytical work, Neon ranks #1 on most Surge AI benchmarks across finance, accounting, spreadsheet modelling and complex constraint following, scores 73.5 percent on Chartography for reading dense professional charts, and performs on par with leading models on Agents' Last Exam v1.1 (75.8 percent). It also takes #1 overall on Zapier's AutomationBench (56.3 percent) and Andon Labs' Vending-Bench 2 ($16,843), a long-horizon simulation of running a business.
Two caveats apply to all of this. First, these are benchmarks Google chose, and vendors always choose the benchmarks that flatter them. Second, benchmark scores measure performance on tests, not performance on your work. The numbers are useful for comparing models against each other, and they are not a promise about what Neon will do with your codebase or your spreadsheets. The independent reporting so far agrees with the broad picture — 9to5Google described the launch as "Carbon quality at Flash speeds" and VentureBeat reported Opus 5.5-level coding at 25 percent lower token prices — but independent testing at scale has not yet had time to happen.
The price: $3 in, $15 out
The pricing is arguably the most important part of the announcement. Neon costs $3 per million input tokens, $0.15 per million cached input tokens — a 95 percent discount on cached input — and $15 per million output tokens, with a 1 million-token context window.
The context window deserves a pause. A million tokens is roughly 750,000 words, or the contents of several long books. In practical terms, it means Neon can hold an entire codebase, a stack of PDFs, spreadsheets, charts, audio and video in a single conversation and synthesise across all of them into structured deliverables. That is the enterprise pitch: not just a chatbot, but a system that can read everything a team has and produce a finished report.
The cost argument is where Google is most aggressive. At its default high thinking level, Google says Neon pairs a high Intelligence Index score with a 32 percent lower cost per task than Carbon, and sits on the Pareto frontier on both the Vals Index and the Agent Arena Index with roughly 50 to 80 percent lower cost per task than top competing frontier models. At the xhigh thinking level, it costs 36 percent or more less per task than competing models at their highest setting. Those are Google's numbers, measured on its chosen benchmarks, but the direction is clear: Neon is priced to undercut the most expensive frontier models while claiming to match much of their capability.
The four thinking levels are part of the same strategy. Each new Gemini 4 model refines the levels, Google says, giving teams "finer control and a smoother progression from one tier to the next" so they can dial intelligence and token spend to what each workload needs. In practice, that means a developer can run routine work at a low thinking level and reserve the expensive xhigh level for the hardest tasks — paying for depth only when it is needed.
Neon for cyber: the security variant
The most technically detailed section of Google's announcement covers a separate variant, Gemini 4 Neon for cyber, which is aimed at defenders rather than general developers.
According to Google, Neon for cyber produces more secure code than Carbon across easy, medium and hard challenges on AutoBaxBench, achieving state of the art on the hard tier at 76 percent. On WebExploitBench v1.2, it is state of the art at 51.26 percent at half the cost of competing large models. On SEC-bench Pro, for finding code vulnerabilities, it reaches 90.5 percent while costing roughly half as much per challenge. On ExploitGym, for validating which vulnerabilities are actually exploitable, it achieves frontier-level performance using half the tokens of Carbon. And on CWE-bench v2, for patching vulnerabilities, it reaches 70.7 percent — a 3.4 percent uplift over Carbon.
The variant is rolling out to trusted defenders in the coming days as part of Google's Fairwind Program, with access requests handled through Cloud account teams. The message to security teams is unmistakable: Google wants its models doing the finding and fixing, not just the writing.
Safety, alignment and the fabrication question
Google devotes a full section of the announcement to safety, and the details are unusually specific. Neon builds on the Gemini 4 safety and security stack, maintaining Carbon-level protections against severe misuse and prompt injection while, Google says, advancing further on alignment and agentic behaviour.
The numbers Google cites: on Gray Swan's indirect prompt injection benchmark, Neon at its highest thinking level resisted every attack, matching a 0.0 percent attack success rate reached only by other Gemini models. It matched Carbon with 0.0 percent cheating on Terminal-Bench 4.0 and took shortcuts on just 2.45 percent of tasks on Google's coding cheating evaluation — the lowest rate of any Gemini 4 model. It achieved a 0.09 percent misaligned action rate on Google's agentic safety suite and a 99.2 percent pass rate on its agent honesty evaluation. It reduced under-service by 43 to 56 percent and overreach by 15 to 35 percent across different effort levels, relative to Carbon's baseline.
On misuse, Google says Neon pairs frontier-level capability with real-time safeguards against cyber and biosecurity misuse comparable to its Carbon protections, while declining far less legitimate work — refusing, for instance, just 1.98 percent of tasks on Terminal-Bench 4.0 and 0 percent on Terminal-Bench-Science.
These are internal Google evaluations, and they should be read as such. But the direction of travel is worth noting: as agentic models get more autonomous, the safety questions shift from "will it say something bad" to "will it do something it should not, or cut corners to finish faster". Neon's reported numbers suggest Google is measuring exactly those behaviours, and that is a meaningful change in what a model launch is expected to prove.
What Neon does in the Gemini app
For consumers, the launch is quieter but still significant. In the Gemini app, Neon is available to Google AI Pro and Ultra subscribers from the model drop-down, and Google says it excels at work that requires frontier-level reasoning with fast responses.
The consumer pitch covers three areas. Higher-quality writing and knowledge work: Neon writes clearly and reads uploads — spreadsheets, PDFs, photos — more precisely, turning dense material into reports and summaries and pulling in details from Google apps like Gmail and Docs. Reliable research and analysis: it frequently incorporates details from sources across the web, for example comparing pricing across vendors before a major purchase or pulling together a complex brief with the latest market data. And faster help for tougher tasks: tasks in Gemini can now run on Neon, working through complex, multi-step jobs more quickly, with usage limits stretching further — analysing billing statements to spot why a utility charge spiked, or cross-checking receipts and documents to negotiate a refund.
As of 19 November, Google had not spelled out Neon's usage caps for each plan, according to AndroidPure's coverage. That is a detail worth watching: a model that is technically available but tightly rate-limited is a different product from one users can actually rely on.
For context on how Gemini's consumer features have evolved, see our explainer on Gemini Personal Intelligence.
The rumours: what Neon is not
The launch has been accompanied by a cloud of speculation about other Gemini 4 variants, and it is worth separating what Google has announced from what it has not.
Hydrogen and Carbon are real. Both are confirmed models in the Gemini 4 family, referenced by name in Google's own Neon announcement, and both have been covered on this site. Hydrogen is Google's fast everyday model; Carbon is its hardest, long-horizon model.
Krypton is not confirmed. The names "Gemini 4 Krypton" and "Gemini 4 Hydrogen" have both trended heavily in search over the past week — "gemini 4 hydrogen" rose more than 3,500 percent in Google Trends queries — but only Neon has a blog.google announcement, and Google has said nothing about a model called Krypton. Treat any claim about Krypton's capabilities, pricing or release date as an unconfirmed rumour until Google announces it.
The same caution applies to the wider family. Google is clearly building a tiered lineup, and more variants are likely. But the lineup that exists today, on the public record, is Carbon, Neon and Hydrogen — plus the cyber variant of Neon rolling out through Fairwind. Anything beyond that is speculation, however confidently it is stated online.
What this means for developers and businesses
Neon's launch lands in the middle of a pricing war that is reshaping how AI is bought. The frontier labs are no longer competing only on who has the smartest model; they are competing on who can deliver near-frontier intelligence at a price that makes it usable for everyday work. Neon's $3 per million input tokens, with cached input at $0.15, is a direct challenge to the premium pricing of the most expensive models, and the 25 percent lower token prices reported by VentureBeat against Opus 5.5-level coding is the kind of comparison that moves procurement decisions.
For developers, the practical questions are straightforward. Is Neon fast enough for agent workflows that run for hours? Google says yes, and points to the Antigravity harness results. Is it accurate enough for production code? The coding benchmarks say yes, but benchmarks are not production. Is it cheap enough to leave running? At $3 per million input tokens, the answer is closer to yes than it has ever been for a model in this class.
For businesses, the 1 million-token context window is the feature most likely to change how work gets done. A model that can hold an entire project — code, documents, spreadsheets, charts — in a single context and produce structured deliverables from it is a different tool from a chatbot that answers one question at a time. Whether Neon lives up to that promise in practice is something only real-world use will establish.
The competitive response will come quickly. Anthropic, which priced its record IPO at $83 a share this week, and OpenAI, which launched GPT-6.2 Astra and Sol days ago, are both shipping at a similar pace. Google's answer, for now, is Neon: frontier-level reasoning, Flash-level speed, and a price that invites developers to use it constantly rather than sparingly.
What We Could Not Verify
- Independent benchmark results. All benchmark figures in this article come from Google's announcement or from outlets reporting it. No independent, large-scale testing of Neon had been published at the time of writing.
- Usage caps. Google had not publicly spelled out Neon's usage limits for AI Pro and Ultra subscribers as of 19 November. Availability in the model picker does not guarantee a specific usage allowance.
- "Gemini 4 Krypton". No announcement, model card or official reference to a model called Krypton could be verified. It remains an unconfirmed rumour.
- Real-world coding performance. Google's coding claims rest on benchmark harnesses (Terminal-Bench, SWE-sweep, FrontierCode). How Neon performs on real production codebases, over long sessions, has not been independently established.
- Regional availability. The announcement describes the rollout in general terms. Availability in specific countries, and any regional restrictions, were not confirmed in the material reviewed.
- Pricing beyond the API. The $3/$15 figures are API prices. Pricing for Neon inside bundled subscriptions (AI Pro, AI Ultra, Gemini for business) was not separately confirmed.
Sources: Google, "Introducing Gemini 4 Neon", blog.google, 18 November 2026; Google Antigravity blog, "Gemini 4 Neon in Google Antigravity"; 9to5Google; VentureBeat; AndroidPure; Artificial Analysis (figures as reported by Google).