Gemini 4 Neon Lands Between Hydrogen and Carbon

Gemini 4 Neon

Google has spent the fourth quarter of 2026 building a family rather than a flagship. Gemini 4 Argon arrived in October for software engineering, Gemini 4 Carbon followed for the hardest long-horizon projects, and Gemini 4 Hydrogen covered fast everyday work at the cheap end. Gemini 4 Neon, announced on 18 November, fills the gap in the middle: a model Google positions for complex coding, analytical reasoning and production workflows that recur often enough that cost per task starts to matter as much as the score on a leaderboard.

That positioning is the interesting part, because it is not a simple story of a smaller model being worse. Gemini 4 Neon scores 61 on the Artificial Analysis Intelligence Index at its highest thinking level, which Google says places it third overall and ahead of its own larger Carbon model at Carbon's highest setting. It also costs more per token than the two mid-tier rivals it is being compared with. The argument Google is making is not that Neon is cheapest. It is that Neon finishes work for less.

This article covers what shipped, how the pricing actually compares, which benchmarks Google chose to publish and which ones it did not lead, why one coding result triples depending on the harness around the model, what the cyber variant does, and how to read the safety claims. Where a number comes from Google, it is attributed to Google. Where it comes from a third-party evaluator, it is attributed there.

What Google shipped

Gemini 4 Neon is available on the paid tier of the Gemini API in Google AI Studio and Android Studio. Google AI Pro and Ultra subscribers can reach it in Google Antigravity, the company's agentic development environment, and Pro and Ultra subscribers can select it in the Gemini app. Enterprises get it through Gemini for business, either by API or from the model drop-down.

The specification is straightforward. Gemini 4 Neon has a one million token context window, which Google frames around mixed inputs rather than raw length: a single request can hold video, audio, charts, spreadsheets and PDFs together. Pricing is 3 dollars per million input tokens, 15 dollars per million output tokens, and 0.15 dollars per million cached input tokens, which Google describes as a 95 percent discount on cache reads.

Compared with the introductory rate of 2 dollars and 10 dollars that Argon and Carbon carried at launch, those headline numbers are 50 percent higher on both input and output. That is the first thing to understand about Gemini 4 Neon: it is not the budget option in the family. Hydrogen is the budget option. Neon is priced above the mid-tier models from rivals and below the flagship tier, and Google's case rests on what a task costs end to end rather than on the sticker price of a token.

Google has also given the model four thinking levels, a structure it says is meant to let teams dial intelligence and spend to the workload instead of choosing a single default. The highest level, xhigh, is where the 61 on the intelligence index comes from. The default is high, and several of the published coding figures are quoted at that default.

Where Gemini 4 Neon sits in the family

The hierarchy Google describes is clean. Hydrogen handles fast everyday execution and wins on raw output speed. Carbon handles the hardest long-horizon work, including multi-day projects and the deepest reasoning. Gemini 4 Neon is positioned for complex work that happens often: coding, analysis and enterprise workflows where a team wants frontier-level capability without paying flagship rates on every run.

That framing explains the pricing. Carbon's introductory rate is below Neon's, which looks odd until the introductory period is taken into account. Google has described the Argon and Carbon rates as introductory, and the standard rates that follow are higher. Once that shift happens, the ordering of the family becomes more intuitive, and the comparison between Gemini 4 Neon and Carbon becomes less about tokens and more about tokens consumed per completed task.

Google's own cost claim is the meaningful one here. At its default high thinking level, Gemini 4 Neon delivers a 32 percent lower cost per task than Carbon on the Artificial Analysis index. At the higher xhigh setting, where the model earns its best intelligence score, Google says the advantage narrows to 36 percent or more below competing models at their highest settings. Those are Google's figures, measured on Google's chosen tasks, and they should be read as the company's argument rather than as an independent finding.

The wider market makes the tier structure harder to read than it used to be. Alibaba's Qwen4-Max launch kept the flagship behind a hosted API while promising smaller open models later, and Anthropic's Fable 5.2 release pushed flagship pricing higher again. In that field a mid-tier model that beats a flagship on an index is a different proposition from a mid-tier model that merely costs less, which is why the harness result later in this article carries more weight than the rank itself.

The benchmarks Google chose

The published results fall into three groups, and the pattern in each is worth noting.

On general capability, Gemini 4 Neon ranks third overall on the Artificial Analysis Intelligence Index at 61, behind larger and slower models. On the Vals Index it ranks second at 70.9 percent, and Google says it reaches first place on the Legal Research Bench and the Tax Agent Bench within that evaluation. On a new Agent Arena Index built from more than 4,200 real-world agent sessions, it debuts second globally with a score of 1,336, and Google separately claims second place on steerability, meaning how cleanly the model adapts to mid-task corrections.

On agentic and analytical work, the wins are more emphatic and more specific. Gemini 4 Neon takes first place on Zapier's AutomationBench at 56.3 percent, and first place on Andon Labs' Vending-Bench 2 with a simulated business result of 16,843 dollars. It scores 73.5 percent on Chartography for reading dense professional charts, 75.8 percent on Agents' Last Exam v1.1, and Google reports it ranks first on most benchmarks in the Surge AI set covering finance, accounting, spreadsheet modelling and complex constraint following.

On scientific and mathematical reasoning, Google claims best in class on Terminal-Bench Science at 70.5 percent, RiemannBench at 84.0 percent and HLE-Diamond at 71.3 percent. Those are strong numbers, and they are also the numbers Google selected. The published set does not include a run against every benchmark where a rival leads, which is normal practice and the reason a buyer should treat a launch post as a starting inventory of tests rather than a verdict.

Where Gemini 4 Neon does not lead

The reporting around the launch has been more enthusiastic than the underlying chart supports, and it is worth being precise about the gaps.

On Terminal-Bench 4.1, Gemini 4 Neon scores 65.2 percent, which is competitive but not first. On FrontierCode 2.0, the evaluation from Cognition that grades code changes against reviewer-set criteria, Google reports 46.2 percent at the default high setting, roughly a point behind the comparison it published. On Chartography, one competitor leads at 75.1 percent against 73.5 percent for Gemini 4 Neon. Those are small margins, and small margins on single benchmarks do not settle which model a team should use.

The larger caveat is configuration. Providers define thinking levels differently, so two models both running at maximum effort can spend very different amounts of time and money on the same question. Google is upfront that its own results move when the harness changes, and the coding section below is the clearest demonstration of why that matters more than any single rank.

The coding result that triples with the harness

The most useful number in the entire launch is not a leaderboard position. It is a pair of numbers from the same model.

On SWE-sweep, an evaluation of software engineering changes, Gemini 4 Neon scores 12.8 percent when run with a minimal agent harness called mini-SWE. In Google Antigravity, with the same model underneath, it scores 39.0 percent, which Google describes as state of the art. The model did not change. The software around it did.

That threefold spread is a more practical insight than any index position. It means a team comparing models on a public leaderboard is measuring the harness as much as the model, and it means the environment a team already runs may be worth more than a marginal capability difference. Google's own framing makes the point explicit: in Antigravity, Gemini 4 Neon can act as an orchestrator directing parallel Hydrogen subagents, dividing a project across several workers and tracking them in real time.

Google also reports that Gemini 4 Neon ranks first on the Artificial Analysis Coding Agent Index at 58.6, which is an agentic measure rather than a raw code generation score. It describes the model as handling multi-file debugging and refactoring with targeted, surgical edits, cutting redundant file reads, sharing progress as it works and compacting context reliably in long sessions. Those are workflow properties rather than capabilities, and they are the ones that determine whether an unattended run finishes or stalls.

Our earlier look at a Gemini 4 Argon coding agent running overnight covers the same question from the opposite direction, and our explainer on the metrics that describe inference speed sets out why throughput figures need care.

Speed, latency and the first token

Gemini 4 Neon is fast once it starts producing output. Launch-day readings place it around 110 tokens per second at the xhigh thinking level, against roughly 59 tokens per second for Carbon at its high setting, which works out to something close to 1.9 times Carbon's output rate. Hydrogen remains the fastest in the family at well over 150 tokens per second depending on setting, and the previous generation's flash model sits near Neon.

The number that matters more for agentic work is the wait before the first token. Independent readings put that at roughly 57 seconds for Gemini 4 Neon at xhigh. A developer watching an agent fix a repository spends that time before any output appears, and then spends more while the agent reads files, runs tests and revises its patch. A fast stream does not compensate for a slow start in a workflow where most of the wall clock goes to tool calls rather than generation.

That distinction is why the harness result and the speed result belong in the same conversation. Output rate describes how quickly text arrives once reasoning is done. Cost per task describes the whole job. For long agent runs, the second number is the one that appears on an invoice.

The cyber variant and the Fairwind Programme

Google is releasing a separate cyber-tuned version of Gemini 4 Neon rather than folding those capabilities into the general model. It is going to trusted defenders through the Fairwind Programme, the same channel that Carbon for cyber used before Carbon's public launch, with access rolling out over the days following the announcement and requests routed through a Google Cloud account team.

The published results for that variant are specific. On AutoBaxBench, which tests whether a model produces secure code, Google claims state of the art on the hard tier at 76 percent, ahead of Carbon across easy, medium and hard challenges. On WebExploitBench v1.2, which tests exploitation, it claims 51.26 percent at roughly half the cost of competing large models. On SEC-bench Pro, which tests vulnerability finding, it claims 90.5 percent, again at about half the cost per challenge.

The patching numbers continue the pattern. On ExploitGym, used to judge which vulnerabilities are actually exploitable, Google says the model reaches frontier-level performance using half the tokens Carbon needs. On Collinear's CWE-bench v2, it reports 70.7 percent, a 3.4 point improvement over Carbon. For security teams, the interesting claim is not the raw score but the ratio: comparable outcomes for roughly half the spend changes how many repositories a team can afford to scan.

Safety claims and how to read them

Google's safety section is unusually quantitative, and it is worth separating what is measured from what is asserted.

The strongest claim concerns prompt injection. On Gray Swan's indirect prompt injection benchmark, where malicious instructions are planted in pages, emails, documents or code that a model reads, Google says Gemini 4 Neon resisted every attack at its highest thinking level with real-time safeguards enabled, matching a zero percent attack success rate that only other Gemini models had reached. That is a benchmark result under a specific configuration, not a guarantee, and Google itself frames prompt injection as an area for continued improvement in its wider documentation.

The behavioural numbers are more granular. Google reports that Gemini 4 Neon refuses 1.98 percent of tasks on Terminal-Bench 4.0 and none on Terminal-Bench Science, which it presents as evidence that stronger safeguards did not make the model unhelpfully cautious. It reports zero cheating on Terminal-Bench 4.0, shortcuts taken on 2.45 percent of tasks in its coding cheating evaluation, a 0.09 percent misaligned action rate on its agentic suite and a 99.2 percent pass rate on its honesty evaluation. It also claims reductions in under-service of 43 to 56 percent and in overreach of 15 to 35 percent compared with Carbon.

Those figures come from Google's internal evaluations and its chosen partners, and none has been independently replicated. The honest reading is that they describe a model Google believes behaves better in long autonomous runs, measured by tests Google designed, which is exactly the kind of claim that a buyer should verify against their own workload before trusting.

Questions readers are asking

What is Gemini 4 Neon? It is Google's mid-tier Gemini 4 model, announced on 18 November 2026 and positioned between Hydrogen for fast everyday work and Carbon for the hardest long-horizon projects. It is built for complex coding, analytical reasoning and production workflows.

How much does Gemini 4 Neon cost? Three dollars per million input tokens, 15 dollars per million output tokens and 0.15 dollars per million cached input tokens. That is 50 percent above the introductory 2 dollar and 10 dollar rates that Argon and Carbon launched with.

How does Gemini 4 Neon compare with Carbon? Google says Gemini 4 Neon scores 61 on the Artificial Analysis Intelligence Index at its highest thinking level, against 60 for Carbon at its high setting, and delivers a 32 percent lower cost per task at its default level. Carbon remains the choice for the longest and hardest projects.

Where can I use it? On the paid Gemini API tier in Google AI Studio and Android Studio, in Google Antigravity for AI Pro and Ultra subscribers, in the Gemini app for Pro and Ultra subscribers, and in Gemini for business for enterprises.

Why did its coding score triple between two tests? Because the harness changed. Gemini 4 Neon scored 12.8 percent on SWE-sweep with a minimal agent harness and 39.0 percent in Google Antigravity with the same model. The surrounding software, not the model, produced most of the difference.

Is the cyber version available? Not generally. A cyber-tuned variant is rolling out to trusted defenders through the Fairwind Programme, with access requested through a Google Cloud account team.

Is Gemini 4 Neon better than its rivals? It leads on several published benchmarks, including Zapier AutomationBench and Terminal-Bench Science, and trails by small margins on others, including Terminal-Bench 4.1 and Chartography. All figures are self-published, so the practical answer depends on a team's own evaluation.

Sources

[fluentform id="2"]