The Fourth Gemini 4 in Seven Weeks: Neon Sits Between Carbon and Hydrogen, Selling Speed With Enough Smarts
On 18 November 2026, Google released Gemini 4 Neon: third overall on the Artificial Analysis Intelligence Index (61) and first on the Coding Agent Index (58.6), priced at $3 per million input tokens and $15 per million output. It sits between the heaviest Carbon and the lightest Hydrogen, conducting teams of Hydrogen subagents in Antigravity — yet per token it is November's priciest Gemini 4, only becoming the cheaper middle option after 1 December.

In short: the fourth Gemini 4 in seven weeks, filling the gap in the middle
On 18 November 2026, Google released Gemini 4 Neon. The official blog post is titled "Introducing Gemini 4 Neon", and its opening line sets the position in one sentence: frontier-level intelligence with fast end-to-end task completion across complex software engineering, knowledge work, and science workflows. The price came out at the same time: $3 per million input tokens and $15 per million output tokens, with cached input at $0.15 per million (a 95% discount) and a 1 million-token context window. It was available from day one: the paid tier of the Gemini API (Google AI Studio and Android Studio), Gemini for business, and for Google AI Pro and Ultra subscribers in Antigravity and the Gemini app.
The name itself reports its position. Google has shipped four Gemini 4 models in seven weeks: Argon on 30 September, Carbon on 28 October, Hydrogen on 12 November, and Neon on 18 November. Carbon is the heavy model for the hardest, longest-running jobs; Hydrogen is the light model for fast everyday work; Neon sits in between. Google's own words are that Carbon tackles your hardest, long-horizon projects, Hydrogen handles fast everyday execution, and Neon balances depth and efficiency, bringing frontier-level reasoning to complex, high-frequency workflows, faster and more cost-efficiently. This article sets out what that "middle" is actually worth, how strong it is, and how it relates to the three models you have already read about.
The report card first: third overall, first at coding
Google lists a lot of benchmarks; three groups say the most about where Neon stands. The first is general intelligence. Neon scores 61 on the Artificial Analysis Intelligence Index at its highest thinking level (xhigh), ranking third overall, behind two larger, slower models. On the Vals Index it ranks second (70.9%) and lands on the Pareto frontier. On Arena's new Agent Arena Index it debuts second globally with a score of 1336, drawn from more than 4,200 real-world agent sessions. On steerability — how cleanly a model adapts when you correct it mid-task — it also ranks second on the Agent Arena Index.
The second group is coding, the part VibeFix readers care about most. Neon ranks first overall on the Artificial Analysis Coding Agent Index with 58.6 points, scores 65.2% on Terminal-Bench 4.1, and reaches 46.2% on FrontierCode 2.0 at its default high thinking level. One comparison deserves to be pulled out on its own: on SWE-sweep, Neon leads the mini-SWE harness at 12.8% and more than triples to a state-of-the-art 39.0% inside Google Antigravity. Same model, different harness, a score more than twice as high — that gap comes back later, because it shows part of Neon's result is supplied by the harness, not by the model working alone.
The third group is enterprise and science work. Neon ranks second on the Vals Index (70.9%), first on both the Legal Research Bench and the Tax Agent Bench, and first on most Surge AI benchmarks across finance, accounting, spreadsheet modeling, and complex constraint following. It scores 73.5% on Chartography for reading dense professional charts, ranks first on Zapier's AutomationBench (56.3%), and first on Andon Labs' Vending-Bench 2 ($16,843), a long-horizon simulation of running a business. On scientific reasoning it is best in class on Terminal-Bench Science (70.5%), RiemannBench (84.0%), and HLE-Diamond (71.3%). The benchmarks have many names, but they point at one selling point: Neon is not a model that only writes code; it is designed to finish a whole job end to end.
Inside Antigravity, Neon is the conductor
Neon has a role no other Gemini 4 model has: inside Google Antigravity, it can act as an Orchestrator, directing a team of parallel Hydrogen subagents. Google describes it concretely: Neon does the judging and the task breakdown while Hydrogen runs the work in parallel; across long tasks Neon cuts down on redundant file reads, shares progress as it works, and handles context compaction reliably. In other words, Google has split "fast" and "smart" across two models: Hydrogen is cheap and quick, Neon costs more and is responsible for thinking and for holding the job together.
That split has a direct meaning for people building with agents: you do not have to choose between "one model does everything" and "switch to a cheaper model to save money". You can let Neon guard the head and tail of a task — breakdown, acceptance, consistency across a long context — and hand the repeated execution to Hydrogen. Google also says Neon handles multi-file debugging and refactoring across repositories with targeted, surgical edits rather than broad rewrites. There is no independent data behind that claim yet, so read it as the vendor describing its own product's behavior, not as a measured result. But the division of labor itself is a real product design: on launch day, Antigravity CLI 1.5.12 added Neon support as well, so API key sessions can call it directly.
Price: expensive first, cheaper later — 1 December is the dividing line
Neon's pricing is the easiest section to misread, so it is worth doing the arithmetic carefully. At list price, Neon costs $3 per million input tokens and $15 per million output tokens; Hydrogen is $1.50 and $7.50, exactly half of Neon; Argon and Carbon cost $2 and $10 during their introductory period, a third less than Neon. Google's claim is that at the default high thinking level, Neon costs 32% less per task than Carbon, and sits on the Pareto frontier on both the Vals Index and the Agent Arena Index at roughly 50% to 80% lower cost per task than top competing frontier models. At xhigh, it still costs 36% or more less per task than competing models at their highest setting.
But "cheaper per task" and "cheaper per token" are two different things. Neon costs more per token than Carbon yet 32% less per task, and the difference comes from finishing the same job with fewer tokens — thinking levels, context compaction, fewer repeated file reads. Read the other way, the same logic holds: if your job does not need Neon's depth of reasoning, Hydrogen does the same work more cheaply, because its list price is half of Neon's.
One date changes this price table: Argon and Carbon introductory pricing ends on 30 November, rising to $4 per million input tokens and $20 per million output tokens from 1 December. From that day, Neon becomes 25% cheaper than both, while Hydrogen remains half its price. In other words, during the twelve days from 18 to 30 November, Neon is the most expensive Gemini 4 model per token; only after 1 December does it become "the cheaper middle option". That timing comes from itdaily's reading of the price table — Google's own blog does not spell it out — so anyone deciding when to switch should check the calendar first.
One more third-party warning is worth including: AI Weekly's hands-on testing found that at the default HIGH thinking level, thinking tokens made up about 90% of the total bill, so the headline $3 and $15 rates are a poor proxy for what a task actually costs. That does not make Google's numbers false; it means the "32% cheaper per task" conclusion depends on the task type and the thinking level. Running one real task of your own and looking at the thinking-token share of the bill is more reliable than doing arithmetic from the price list.
Four thinking levels: Neon turns "how long to think" into a dial
Neon ships with four thinking levels. Google says it refines the levels with each new model release, 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. The levels scale steadily rather than switching on and off: easy work takes a low level and saves tokens; hard work takes a higher level and thinks longer. Neon's highest level, xhigh, is where its Intelligence Index score of 61 comes from, and where its Agent Arena score of 1336 was earned; the default high level is where the FrontierCode 2.0 score of 46.2% and the "32% cheaper per task than Carbon" claim come from. Same model, different levels, two different sets of numbers — check which level a benchmark used before quoting it.
For everyday use, the level is the first valve on cost, not a decoration. You used to control cost by switching models; now you can control "how deep to think" inside one model. Low levels suit work with a fixed format and clear steps, such as generating a report from a template or editing code from a checklist; high levels suit work that needs multi-step reasoning, such as tracing a performance regression across repositories or assembling an accurate brief from several contradictory documents. Setting the level wrong costs in both directions: too low, and the model takes shortcuts and fabricates results on complex tasks — Google itself says Neon is still improving on completing multi-step work without cutting corners, fabricating results, or overstepping boundaries; too high, and thinking tokens push the bill up, with AI Weekly's testing putting them at about 90% of the total at the default HIGH level.
One more thing worth noting: the level and the price move together. Google says xhigh still costs 36% or more less per task than competing models at their highest setting — but the comparison is against "the competitors' highest setting", not against Neon's own low level. Moving from Neon's low level to xhigh will raise your own bill, by an amount that depends on the task. A steady approach is to run a week of real tasks at the default high level, note the thinking-token share of the bill and the task success rate, and only then decide which tasks deserve a higher level and which can drop lower. The level is a dial, not a switch; anyone who sets it once and forgets it leaves savings on the table.
The enterprise and science numbers, taken apart
The enterprise and science benchmarks deserve a closer look, because they answer a concrete question: besides writing code, can Neon actually finish a real job? Start with enterprise work. Neon ranks second on the Vals Index (70.9%), which measures model performance on real business tasks rather than multiple-choice questions; it ranks first on both the Legal Research Bench and the Tax Agent Bench, which test whether a model can give defensible answers in a field where answers have right and wrong and a legal basis. On Surge AI benchmarks it ranks first on most categories across finance, accounting, spreadsheet modeling, and complex constraint following — work where one wrong digit means one wrong report. Chartography is 73.5% for reading dense professional charts, Zapier's AutomationBench is 56.3% for first place, and Vending-Bench 2 comes in at $16,843 for first place, measuring long-horizon, multi-step business decisions with feedback, not single-turn answers.
The scientific reasoning numbers are not decoration either. Terminal-Bench Science at 70.5%, RiemannBench at 84.0%, and HLE-Diamond at 71.3% are all described by Google as best in class. Terminal-Bench Science tests completing scientific computing tasks in a terminal, RiemannBench tests mathematical reasoning, and HLE-Diamond tests hard general reasoning. What they share is that they measure not "how much the model knows" but "whether it can walk a multi-step chain of reasoning to the end without breaking". For people doing vibe coding, the practical value is this: if you ask a model to read papers, run data analysis, or find bugs in scientific computing code, Neon's long reasoning is where it pulls away from Hydrogen — Hydrogen is fast, but the reasoning that has to be walked to the end is exactly what Neon's extra cost buys.
The sourcing of these numbers needs to be stated plainly, though. Every score above comes from Google's own blog. The leaderboards are third-party, but the statement "Neon ranks this high with this score" is Google's own reporting, with no independent third party having reproduced the same set of scores yet. Vals, Artificial Analysis, and Arena are public leaderboards whose scores can be checked, but figures tied to a thinking level — such as "FrontierCode 2.0 is 46.2% at the default high level" — exist only in Google's measurement. Treat the rankings as a position reference and the exact scores as vendor-reported results: the positions are fairly credible, the scores await independent reproduction.
The cyber section: Neon for cyber is a product line, not a security incident
The blog gives cybersecurity its own section. Neon for cyber is a variant for trusted defenders, rolling out over the coming days through the Fairwind Program. On the numbers, Neon reaches state of the art on the hard tier of AutoBaxBench (76%); state of the art on WebExploitBench v1.2 (51.26%) at half the cost of competing large models; state of the art on SEC-bench Pro (90.5%) at roughly half the cost per challenge; frontier-level performance on ExploitGym using half of Carbon's tokens; and 70.7% on CWE-bench v2, a 3.4% uplift over Carbon.
To be clear about what this section is: it describes "a model that helps defenders do their work", not "a security incident happened". Neon producing more secure code than Carbon across easy, medium, and hard AutoBaxBench challenges is a product capability; the Fairwind Program is a targeted rollout, not a vulnerability disclosure. VibeFix has run security stories in three straight cycles; the cybersecurity content here is part of a model launch, so the theme remains a new model release and it does not count as a security story — the same treatment cycle226 gave to the AES-256-GCM encryption content in the Vercel piece.
The blog's safety commitments carry another set of numbers, and it matters which are third-party benchmarks and which are Google's own internal evaluations. Neon keeps Carbon-level protections against severe misuse and prompt injection, refusing just 1.98% of tasks on Terminal-Bench 4.0 and 0% on Terminal-Bench Science. It matched Carbon with 0.0% cheating on Terminal-Bench 4.0, took shortcuts on just 2.45% of tasks on Google's coding cheating evaluation (the lowest rate of any Gemini 4 model), recorded a 0.09% misaligned action rate on its agentic safety suite, and a 99.2% pass rate on its agent honesty evaluation. On Gray Swan's indirect prompt injection benchmark, with real-time safeguards at its highest thinking level, it resisted every attack — a 0.0% attack success rate. The benchmark names are third-party, but the scores are Google's own reporting — like every benchmark figure in this article, they have no independent reproduction yet.
In the Gemini app, Neon is for the awkward errands
Neon is not only for developers and enterprises; Google AI Pro and Ultra subscribers can pick it directly in the Gemini app. Google's consumer positioning is "work that requires frontier-level reasoning and fast responses": reading uploaded spreadsheets, PDFs, and photos more precisely, turning dense material into well-structured reports and summaries, and pulling relevant details from Google apps like Gmail and Docs; keeping research current by bringing in details from across the web, such as comparing prices across vendors before a major purchase or pulling together a complex brief with the latest market data; and working through tedious multi-step to-dos more quickly while stretching usage limits further, like analyzing billing statements to spot why a utility charge spiked, or cross-checking receipts and documents to negotiate a refund.
Those descriptions sound like everyday errands, but they are the same job as the developer use case: multi-step, context-heavy, and not allowed to break halfway. Working out why a bill went up and debugging a cross-repository performance regression use the same capability — pulling scattered information together, walking a multi-step chain of reasoning, and producing a conclusion that can be checked. Neon's consumer selling point is putting that capability inside everyone's app, not only in the API and Antigravity. For subscribers, the practical change is that the errands that used to feel too slow for a big model and too fabrication-prone for a small one now have a middle option. Whether that is worth a Pro or Ultra subscription depends on how many such errands you actually have — if you hit them twice a month, the default model is enough.
Boundary with what VibeFix already published: this is a fourth model, not a sequel
VibeFix has already published four Gemini 4 articles, so it is easy to mistake Neon for a follow-up to one of them; the boundary is worth drawing once. The first was Argon (released 30 September, published on VibeFix in early October), the family's first model, aimed at complex, long-running tasks. The second was Carbon (released 28 October), the heaviest model in the family. The third was Carbon's general availability (10 November), which covered Carbon moving from staged rollout to full availability — not a new model. The fourth was Hydrogen (released 12 November), the light, fast model with four thinking levels, 2.5 times faster than the previous generation.
Neon's relationship to those four is that it is a new reporting subject released on 18 November — the fourth model in the family, with its own price, its own benchmark results, and its own rollout. It is not a Carbon availability update (that was the third article), not a Hydrogen rename (the price and the positioning both differ), and not a repeat of Argon. The one point of possible confusion is the Orchestrator role in Antigravity: the Hydrogen article also mentioned Antigravity, but there Hydrogen was the execution layer; here Neon is the conductor and Hydrogen is the subagent, the roles reversed. Do not mix the two articles' roles up.
For people building with agents: when to pick Neon, and when not to
Translated into everyday choices, the answer depends on whether your task is "hard" or "frequent". If the work is multi-file debugging, cross-repository refactoring, or long tasks with many tool calls, where a break is expensive because re-running burns tokens and re-doing tool calls, Neon is the model in the family priced for that work: the thinking level can be tuned, context compaction is described by Google as reliable, and fewer redundant file reads happen on long tasks. If the work is everyday, repetitive, and done in minutes, Hydrogen is cheaper at half Neon's list price, and there is no reason to pay for reasoning depth you will not use. If the work is the hardest, hours-long project in the building, Carbon remains the heaviest model in Google's positioning; Neon does not replace it.
Two practical reminders. First, Neon's "32% cheaper per task" is calculated at the default high level against Carbon; change the thinking level or the task type and the number moves. Before switching, run one real task of your own and look at the thinking-token share of the bill, then decide which level to use by default. Second, the 1 December price change affects long-run costs: if your usage extends past December, any cost estimate made on November prices needs to be redone on 1 December — Neon will be cheaper than Carbon then, Hydrogen will still be the cheapest, and the ranking inside the family will shift.
Conclusion: model launches are getting denser, and the middle tier is the new normal
Four models in seven weeks pushes Google's Gemini 4 release pace to a new speed: Argon, Carbon, Hydrogen, Neon, each finding its slot in the gaps left by the last. Neon's slot says something about where the model race is heading: the contest is no longer only about "who is smartest", but about "who can finish a job at a given price". An Intelligence Index score of 61, first place on the Coding Agent Index at 58.6, a 32% cost-per-task advantage — each number looks bright on its own, but what matters is what they add up to: frontier models are becoming things you pick by task, by thinking level, and by the bill.
For people doing vibe coding, that change has a practical meaning: choosing a model is increasingly like choosing a tool — not picking "the strongest", but picking "the best value for this task". Neon adds a middle answer to that question: no need to pay Carbon prices for the hardest jobs, no need to settle for Hydrogen's lightness on complex ones. The arrival of a middle tier is itself the signal that models are starting to be organized around real workflows. Which slot the next model fills, and how the price table looks after 1 December, is worth watching.
Comments (0)
Sources
Related articles
On 17 November 2026, Microsoft launched MAI Code 2 and Code 2 Flash: the larger model has 77 billion active parameters and a one-million-token context, averaging 63.2% across four benchmarks — close to GPT-6.1 Sol and Claude Opus 5.5 — at $0.309 per task. Read the three caveats first: promotional pricing, the lowest score of the three, and no long-horizon work.
On 17 November 2026, Perplexity launched the Orchestrator API: one interface reaching OpenAI, Anthropic, Google, xAI and open-weight models such as GLM, Kimi and DeepSeek, billed at the answering model's published rate with no platform fee. When one provider fails, the same model retries on another deployment, and preview routing assigns requests to models by task.
On 11 November 2026, AWS open-sourced UAEF: adapters turn traces from LangGraph, LangChain, Strands and Bedrock AgentCore into one format, then 36 metrics across 7 dimensions score them, flagging regressions past 5%. The judge model is hardened too — scores must cite verbatim evidence, and never decide safety.