Google has launched a fourth Gemini 4 model in five weeks. Gemini 4 Neon arrived on November 18, and it is the clearest sign yet of where the AI race is heading. Not toward the smartest model money can buy. Toward the smartest model money can afford to run all day.
Neon sits between the two models Google released earlier this month. Carbon handles the hardest, long-horizon projects. Hydrogen handles fast everyday work. Neon is built for the complex, high-frequency work in the middle: long coding sessions, enterprise analysis, and science workflows that need frontier-level reasoning without frontier-level bills.
The price tells the story. Neon costs 3 dollars per million input tokens and 15 dollars per million output tokens, with cached input at 15 cents, a 95 percent discount. It carries a 1 million-token context window. At its default high thinking level, Google says Neon costs 32 percent less per completed task than Carbon. Against top competing frontier models, Google puts the saving at roughly 50 to 80 percent per task.
That is the pitch. Whether it holds up in real use is the question every builder should now test.
What Neon is
Neon is a frontier model tuned for end-to-end task completion. Google describes it as bringing frontier-level intelligence with fast completion across complex software engineering, knowledge work, and science workflows. It reads spreadsheets, PDFs, charts, audio, and video inside a single 1 million-token context window and turns them into structured deliverables.
It has four configurable thinking levels, so teams can dial intelligence and token spend up or down per workload. The default is high. The top level, xhigh, adds deeper multi-step reasoning for the hardest engineering and analytical work. Google says even at xhigh, Neon still costs 36 percent or more less per task than competing models at their highest settings.
Availability is paid only. Neon is rolling out in the Gemini API paid tier in Google AI Studio and Android Studio, in Gemini for business, in Google Antigravity for paid Google AI Pro and Ultra subscribers, and in the Gemini app for Pro and Ultra subscribers. There is no free tier. In the Gemini app, Neon replaces the Argon model in the picker, and the older Gemini 3.1 Pro is being removed, according to 9to5Google.
The model card says Neon is adapted from Carbon, with a knowledge cutoff of March 2026. That matters for anyone using it on fast-moving topics. It will not know what happened this autumn unless it looks it up.
The numbers Google is proud of
Every benchmark below comes from Google, and Google chose which ones to publish. Treat them as a claim to verify, not as settled fact. Still, the pattern is worth reading, because it shows what Google thinks Neon is for.
Neon ranks third overall on the Artificial Analysis Intelligence Index with a score of 61 at its xhigh thinking level, behind larger and slower models. On the Vals Index it ranks second at 70.9 percent, and it takes first place on both the Legal Research Bench and the Tax Agent Bench. It ranks second on steerability, which measures how cleanly a model adapts when a user corrects it mid-task, on the new Agent Arena Index, where it debuts second globally with a score of 1336 across more than 4,200 real-world agent sessions.
On coding, Neon ranks first overall on the Artificial Analysis Coding Agent Index at 58.6. It 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 the mini-SWE harness at 12.8 percent and reaches 39.0 percent inside Google Antigravity, which Google presents as evidence that its own harness unlocks more of the model.
On science and maths, Neon is best in class on Terminal-Bench Science at 70.5 percent, RiemannBench at 84.0 percent, and HLE-Diamond at 71.3 percent. It scores 75.8 percent on Agents’ Last Exam v1.1 and 73.5 percent on Chartography, a test of reading dense professional charts.
On business workflows, Neon ranks first on Zapier’s AutomationBench at 56.3 percent and first on Andon Labs’ Vending-Bench 2, a long-horizon simulation of running a business, with a score of 16,843 dollars. It also ranks first on most Surge AI benchmarks across finance, accounting, spreadsheet modeling, and complex constraint following.
The through line is consistent. Neon is not trying to win the single hardest benchmark. It is trying to win the long, multi-step jobs that make up real work: the coding session, the research brief, the spreadsheet model, the business simulation.
Built for code, and for cyber defence
Google is pitching Neon hard at developers. It is engineered for long-running agent workflows and sustains performance through complex, tool-heavy tasks end to end. In Antigravity, it can act as a fast, high-judgment Orchestrator directing parallel Hydrogen subagents in a single session. It cuts down on redundant file reads, shares progress as it works, and handles context compaction reliably in long sessions. For production code, Google says it handles multi-file debugging and refactoring across repositories with targeted, surgical edits.
There is also a separate cyber variant. Gemini 4 Neon for cyber produces more secure code than Carbon across easy, medium, and hard challenges on AutoBaxBench, reaching state of the art on the hard tier at 76 percent. On WebExploitBench v1.2, a penetration testing benchmark, 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 at roughly half the cost per challenge. On ExploitGym, for validating which vulnerabilities are actually exploitable, it reaches frontier-level performance using half the tokens of Carbon. On CWE-bench v2, for patching, it reaches 70.7 percent, a 3.4 percent uplift over Carbon.
The cyber variant is rolling out to trusted defenders in the coming days as part of the Fairwind Program. That gated rollout follows the same pattern Google used for Carbon, which went to defenders first while safeguards were tested.
GitHub has also made Neon generally available in Copilot, across Pro+, Max, Business, and Enterprise plans, in VS Code, JetBrains IDEs, and its other surfaces. That is a meaningful distribution win. A model that shows up where developers already work gets used. A model that only lives in a demo does not.
Safety, without the usual trade-offs
Faster and cheaper models usually come with a safety catch. Google’s argument is that Neon avoids one. It says Neon maintains Carbon-level protections against severe misuse and prompt injection while refusing far less legitimate work: just 1.98 percent of tasks on Terminal-Bench 4.0 and 0 percent on Terminal-Bench-Science.
The alignment numbers Google reports are strong, and like the benchmarks, they are Google’s own. Neon 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 recorded 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. Under-service fell by 43 to 56 percent and overreach by 15 to 35 percent across different effort levels, against Carbon’s baseline.
On Gray Swan’s indirect prompt injection benchmark, where malicious instructions are planted in web pages, emails, documents, or code the model accesses, Neon resisted every attack at its highest thinking level, a 0.0 percent attack success rate. Google also says Neon records its lowest fabrication rate yet in internal Antigravity testing.
None of this is independently verified yet, and the full Safety and Security Report for Neon is still marked as coming soon. But the direction is clear. Google is trying to prove that a cheaper, faster frontier model can also be a more disciplined one, and it is publishing the measurements to be judged on.
Why the middle tier is the real product
Step back from the benchmark table and the launch itself is the story. Google has now shipped four Gemini 4 models in five weeks: Argon on October 13, Carbon on October 28, Hydrogen on November 12, and Neon on November 18. No other lab is releasing frontier models at this pace. OpenAI shipped GPT-6.2 Astra and Sol on the same day this week. Anthropic shipped Claude Fable 5.2 and Mythos 5.2 last week. The frontier race has become a release cadence race, and Google is setting the tempo.
But Neon also reveals a shift in what the race is about. The first phase was capability: who has the smartest model. The second phase is cost per completed task. Carbon proved Google could build a frontier model. Hydrogen proved it could build a cheap, fast one. Neon is the model that tries to be both at once, and the 32 percent saving per task against Carbon is the number Google leads with, not the benchmark rank.
That shift favours Google more than any other lab. It owns the distribution: Search, Workspace, Android, the Gemini app, and now GitHub Copilot. A model that is nearly as smart as the best and far cheaper to run is the one that ends up inside products people already use. The most expensive model in the world is a trophy. The cheapest useful one is infrastructure.
There is also a pricing clock ticking. Carbon’s introductory pricing ends on November 30, and it rises to 4 dollars per million input tokens and 20 dollars per million output tokens from December 1, according to 9to5Google. At that point Neon, at 3 and 15 dollars, becomes roughly 25 percent cheaper than Carbon on list price as well as cheaper per task. Google is not just launching a new model. It is steering its own customers toward the cheaper one.
What builders should do
First, test Neon on your own longest job, not on a benchmark. The model’s case rests on sustained performance through multi-step work, so give it a real task: a multi-file refactor, a research brief from a pile of PDFs, a spreadsheet model with messy source data. Measure the cost per completed task, not the cost per token. That is the metric Google is asking to be judged on, and it is the only one that matters to your bill.
Second, try the Orchestrator pattern in Antigravity. Neon as planner with Hydrogen as executor is Google’s own recommended setup, and it mirrors how multi-agent work is actually evolving: a smart model that plans and checks, and cheap models that do the bulk work. If your agent stack is one model doing everything, it is probably overpaying.
Third, watch the cyber variant if you ship code. A model that writes more secure code from the start, finds vulnerabilities, validates them, and patches them is a different kind of tool from a model that just writes code fast. The Fairwind rollout is gated, so most teams will wait, but the direction is worth planning around: security review is becoming part of the model, not a step after it.
Fourth, do not build on the assumption that Neon is the last word. Google has shipped four models in five weeks, and the family is clearly not finished. Treat the model as a replaceable part. Keep your prompts, your evaluation set, and your cost tracking outside any single model, so the next launch is a swap, not a rewrite.
What to watch
The first independent benchmarks will matter more than Google’s launch numbers. Artificial Analysis and other third-party leaderboards will show whether Neon’s third-place rank and its cost-per-task claims survive outside Google’s harness. Early testers have already started putting it through coding work, and their verdicts will land over the next few days.
The second thing to watch is uptake. Neon is paid only, and it replaces Argon in the Gemini app. How many Pro and Ultra subscribers switch, and how many developers pick it in the API over Carbon and Hydrogen, will tell us whether the middle tier is where the market actually wants to be.
The third is price. Carbon’s step up on December 1 will make Neon’s discount concrete. If Google follows with more price moves, or if OpenAI and Anthropic answer with their own mid-tier pricing, the cost race will accelerate. That is good for builders and bad for anyone whose business model depends on intelligence staying expensive.
Google has spent five weeks proving it can ship frontier models quickly. With Neon, it is now proving it can ship one cheaply. The race is no longer only about who builds the smartest model. It is about who makes frontier intelligence affordable enough to use all day, every day. Neon is Google’s answer, and for now it is a strong one.
Related reading: Google’s New Gemini Is Fast, Cheap, and Open to Everyone | Google Built Gemini 4 Carbon to Fight Hackers | Google Just Released Gemini 4 Argon. Here Is What It Changes
make it pop
Share card studio
More from Signals
All in topicAnthropic Opened at $2 Trillion. It Closed at $97 a Share.
Anthropic is now a public company, and the market has put its first price on a frontier AI lab. The maker of Claude opened on the…
Anthropic Priced Its IPO at $83. The Market Just Valued Claude at $1.5 Trillion
Anthropic priced its initial public offering at $83 a share on Wednesday, above the $70 to $78 range it set last week. The final price was…
OpenAI’s GPT-6.2 Is Cheaper, Sharper, and Built for Real Work
OpenAI shipped two new models on November 17, and the most interesting thing about them is not how smart they are. It is how cheap they…
Nvidia Just Doubled Its Revenue Without Jensen Huang
Nvidia reported $115.8 billion in quarterly revenue, more than double a year ago, in its first earnings report since Jensen Huang resigned. The business is stronger…
Keep exploring