Friday, 20 November 2026
NEXA NEWSHD
Artificial Intelligence

Google Launches Gemini 4 Neon, Its Fourth Model in Seven Weeks, Promising Carbon-Level Power at Flash Speed

Google has launched Gemini 4 Neon, a new artificial intelligence model it says delivers close to the reasoning power of its flagship Gemini 4 Carbon while running far faster and costing less per task, in the latest salvo of a model-release race that has now produced four new Gemini models in about seven weeks.

Google announced Neon on Tuesday, 18 November, in a blog post by Koray Kavukcuoglu, senior vice president at Google DeepMind and Google’s chief AI architect, according to Google. The model began rolling out the same day through the Gemini application programming interface (API), Google AI Studio, Gemini for business and Google Antigravity, and to paying Google AI Pro and Ultra subscribers in the Gemini app.

Neon sits in the middle of the Gemini 4 family, between the fast, everyday Hydrogen model and the heavyweight Carbon model built for the longest, hardest projects. Google describes it as bringing “frontier-level intelligence with fast end-to-end task completion across complex software engineering, knowledge work, and science workflows”.

The launch lands in the middle of one of the busiest weeks in the short history of the AI industry. OpenAI released its GPT-6.2 Astra and Sol models on Monday, 17 November, Anthropic priced the largest initial public offering in United States history at $83 a share on Wednesday, 18 November, with trading due to begin on Thursday, and Nvidia reported a record $115.8 billion quarter the same week while halting licensed chip shipments to China. Google, which SiliconANGLE noted had “laid low over the summer while Anthropic and OpenAI released new frontier models”, has now answered with a model a week.

What Gemini 4 Neon is

Neon is, in Google’s own description, a model “adapted from Gemini 4 Carbon”, according to its model card, published alongside the launch. It is built for complex coding, analytical reasoning and production-grade enterprise work, and it is aimed squarely at the high-frequency workflows — the coding sessions, document analyses and agent tasks that developers and businesses run again and again, rather than the one-off flagship demonstrations that grab headlines.

The model accepts text, images, audio, video and PDF files as input and replies in text only, according to Google’s documentation. It does not generate images or audio, and it does not support the Live API, which means it is a workhorse for agents and long tasks rather than a media model. Its context window — the amount of material it can hold in a single request and response combined — is one million tokens, the same as Hydrogen, and its knowledge runs to March 2026, with Google warning that some domains may be less current.

A standard synchronous response can run to 64,000 tokens by default, while a background “long decode” mode can stretch output up to the full one million tokens, according to the model card. That matters for the long, multi-step jobs Neon is designed for: a model that can keep working through a large codebase or a stack of documents without losing the thread.

Neon also introduces a new discipline to Google’s pricing of intelligence. It has four “thinking levels” — low, medium, high and extra high, or xhigh — which control how long the model reasons before it answers and therefore how many tokens a user pays for. Google says the levels scale steadily, letting teams “dial intelligence and token spend to what each workload needs”. At the default high level, Neon pairs a high benchmark score with a cost per task 32 per cent lower than Carbon, according to Google; at xhigh it delivers its highest intelligence score while still costing, Google says, 36 per cent or more less per task than competing models at their highest settings.

The launch reshuffles the Gemini app itself. According to 9to5Google, Neon replaces the Argon model in the app’s model picker, and the long-running Gemini 3.1 Pro is being retired. Paid users now see four choices: Auto, which for paid users runs on Gemini 3.5 Flash-Lite; Everyday help, which is Hydrogen; Deeper analysis, which is Neon; and Code and complex work, which is Carbon.

The price — and the awkward promotion

Through the paid Gemini API, Neon costs $3 per million input tokens and $15 per million output tokens, with cached input — text the model has already processed and can reuse — at $0.15 per million tokens, a 95 per cent discount, according to Google. The Batch API, for work that does not need an immediate answer, costs exactly half the standard rate across the board.

On paper, that makes Neon 25 per cent cheaper per token than Carbon’s standard rate. But the paper has a wrinkle, and it is one Google’s own pricing page makes plain. Carbon — along with Argon — is currently on an introductory price of $2 per million input tokens and $10 per million output tokens, which runs until 30 November. Until that promotion ends, the flagship is actually cheaper than the mid-tier model. Only from 1 December, when Carbon’s price doubles to $4 and $20, does Neon become the cheaper option, at 25 per cent below Carbon’s standard rate. Hydrogen, the everyday model, costs $1.50 and $7.50 — half of Neon’s price.

The inversion is a small thing, and it lasts twelve days. But it captures something real about the moment: Google is repricing its whole ladder in public, in the middle of a release sprint, and the numbers on the pricing page are moving faster than the marketing. For a developer choosing a model this week, the rational choice on price alone is Carbon — the most powerful model in the family, at a promotional discount — until the last day of November. From December, the arithmetic flips, and Neon becomes the value pick. Google’s bet is that most buyers will not switch models every fortnight, and that the model they settle on for December is the one that matters.

There is no free tier for Neon. Free Gemini users continue to use Hydrogen, and Neon is available only to paying subscribers and API customers, according to Google and 9to5Google. That is itself a statement: the frontier, or something close to it, is now a paid product, and Google is drawing a clear line between the free everyday tier and the models it charges for.

The benchmarks Google is trumpeting

Google’s announcement is, as ever, built around a long list of test results, and the headline numbers are striking. On the Artificial Analysis Intelligence Index, Neon scores 61 at its highest thinking level, ranking third overall — just behind larger, slower models, according to Google. On the Vals Index, which measures professional tasks, it ranks second with a score of 70.9 per cent, and reaches first place on both the Legal Research Bench and the Tax Agent Bench. It debuts second globally with a score of 1,336 on Arena’s new Agent Arena Index, across more than 4,200 real-world agent sessions.

In coding, Google says Neon ranks first overall on the Artificial Analysis Coding Agent Index with a score of 58.6, and scores 65.2 per cent on Terminal-Bench 4.1 and 46.2 per cent at its default high setting on FrontierCode 2.0. On SWE-sweep, a software-engineering benchmark, it leads on the mini-SWE harness at 12.8 per cent and more than triples to a state-of-the-art 39.0 per cent inside Google Antigravity, which Google presents as evidence of what its own agent harness unlocks.

In science and mathematics, Google presents Neon as best in class on Terminal-Bench Science at 70.5 per cent, RiemannBench at 84.0 per cent and HLE-Diamond at 71.3 per cent. On the long-context GraphWalks test it scores 95.2 per cent, against 66.8 per cent for Anthropic’s Claude Opus 5.5 and 80.1 per cent for OpenAI’s GPT-6.2 Sol, according to coverage of the launch. It also ranks first on Zapier’s AutomationBench at 56.3 per cent and on Andon Labs’ Vending-Bench 2, a long-horizon simulation of running a business, with a score of $16,843. On Surge AI’s evaluations it ranks first on most benchmarks across finance, accounting, spreadsheet modelling and complex constraint-following, and it scores 73.5 per cent on Chartography, for reading dense professional charts.

Google’s model card compares Neon against Claude Opus 5.5 and GPT-6.2 Sol across 19 rows, and according to ITmedia’s reading of the card, Neon comes first in 16 of them and behind in three — including Terminal-Bench 4.1, where Opus 5.5 beats it by a few tenths of a point. The two models that sit ahead of Neon on the Artificial Analysis index — Anthropic’s Claude Fable 5.2, with 65 points, and OpenAI’s GPT-6.2 Astra, with 62, according to SiliconANGLE — do not appear in Google’s comparison table at all. Google calls Neon “competitive with the leading models in its class”, which is, as one analysis noted, a carefully drawn class.

The independent numbers — and where they differ

Every benchmark in a launch announcement deserves the same health warning, and this one is no exception: most of the numbers Google published are Google’s own claims, on tests Google chose to highlight, and nobody outside the company had independently run them when the model launched. That does not make them false. It makes them a manufacturer’s specification sheet, not a review — and the early independent readings that do exist tell a slightly more modest story than the press release.

The one independently confirmed ranking comes from Artificial Analysis, the independent benchmarking firm, which measured Neon at 61 on its Intelligence Index at the xhigh setting, one point above Carbon at high effort and third overall, behind Claude Fable 5.2 at 65 and GPT-6.2 Astra at 62, according to SiliconANGLE. That confirms the headline placement. But the same firm’s measurements put the speed and cost advantage in a different light from Google’s claims. Google says Neon completes most tasks about three times as quickly as Carbon at costs about 60 per cent lower, according to SiliconANGLE’s account of the announcement. Artificial Analysis, by contrast, measured roughly 110 output tokens per second for Neon against about 57 for Carbon — nearly twice as fast, not three times — and $6.44 per task at xhigh against $7.76 for Carbon, roughly 17 per cent less, not 60 per cent, according to The Decoder’s reading of the firm’s data. A launch-day tracker cited by The Outpost put Neon about 1.9 times faster than Carbon on output rate, closer to Artificial Analysis than to Google’s own figure.

The gap is not a contradiction so much as a difference in what is being measured. Google’s three-times and 60-per-cent figures come from its own internal evaluations of end-to-end task completion, where Neon’s efficiency — fewer redundant file reads, better context handling, faster first response — compounds. Artificial Analysis measures output tokens per second and the cost of completing its own index tasks. Both can be true at once, and both are partial. What matters for a buyer is that the independent numbers confirm the direction of Google’s claim — faster, cheaper, nearly as smart — while trimming its size.

There is also an early independent test from Spring Prompt, a small benchmarking outfit, which ran Neon through five business benchmarks on the day of release and found it set a new best on factual product listings, with 78.6 per cent of its listings ready to publish in every run — ahead of all 35 models it has measured, including GPT-6.2 Sol, GPT-6.2 Astra and Claude Fable 5.2 — while proving weaker on sales copy, slide decks and survey analysis. One small test proves little on its own, but it hints at the shape of Neon’s strengths: precise, factual, structured work, rather than creative flair.

Built for code, agents and cyber defence

The heart of Google’s pitch is not general chat but production work: long-running agent workflows, multi-file debugging and refactoring across repositories, and the orchestration of parallel subagents. Neon is engineered, Google says, to sustain performance through complex, tool-heavy tasks end to end, whether running standalone or acting as a fast, high-judgment orchestrator directing parallel Hydrogen subagents inside Antigravity. It cuts down on redundant file reads, shares progress as it works, and handles context compaction reliably in long sessions — the unglamorous plumbing that decides whether an agent finishes a job or falls apart halfway through.

Google also makes an unusually detailed case for Neon as a cybersecurity tool, and it is here that the launch is most interesting. As agentic coding scales, Google argues, generating secure code from the start matters more than patching it afterwards. On AutoBaxBench, Neon produces more secure code than Carbon across all easy, medium and hard challenges, achieving state-of-the-art performance on the hard tier at 76 per cent, according to Google. On WebExploitBench version 1.2, a penetration-testing benchmark, it is state of the art at 51.26 per cent, at half the cost of competing large models. On SEC-bench Pro, for finding code vulnerabilities, it reaches 90.5 per cent 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 version 2, for patching vulnerabilities, it reaches 70.7 per cent, a 3.4 per cent uplift over Carbon.

Neon for cyber is rolling out to trusted defenders over the coming days as part of Google’s Fairwind Program, according to Google. The pitch is deliberate: the same model that writes code can also be the model that breaks into systems to test them, and Google wants to sell both sides of that work to the defenders first.

For enterprise work more broadly, Google says Neon can synthesise across spreadsheets, PDFs, charts, audio and video inside a single one-million-token context window and turn them into structured deliverables — the kind of work that lives in a finance department or a law firm rather than a chatbot window. Its first-place rankings on the Legal Research Bench and the Tax Agent Bench, and on Surge AI’s finance and accounting evaluations, are aimed squarely at that audience.

Safety: strong claims, incomplete paperwork

Google’s safety claims for Neon are confident, but the paperwork behind them is, by the company’s own admission, unfinished. The model builds on the Gemini 4 safety and security stack, maintaining what Google calls Carbon-level protections against severe misuse and prompt injection, while advancing further on alignment and agentic behaviour. It refuses just 1.98 per cent of tasks on Terminal-Bench 4.0 and none on Terminal-Bench Science, which Google presents as evidence that it declines far less legitimate work than its predecessors. It matched Carbon with 0.0 per cent cheating on Terminal-Bench 4.0 and took shortcuts on just 2.45 per cent of tasks on Google’s coding-cheating evaluation, the lowest rate of any Gemini 4 model. It achieved a 0.09 per cent misaligned-action rate on Google’s agentic safety suite and a 99.2 per cent pass rate on its agent honesty evaluation, and it reduced under-service by 43 to 56 per cent and overreach by 15 to 35 per cent across different effort levels, according to Google.

On indirect prompt injection — malicious instructions planted in web pages, emails, documents or code the model accesses — Google says Neon resisted every attack in Gray Swan’s IPI benchmark at its highest thinking level, matching a 0.0 per cent attack success rate reached only by other Gemini models.

But the model card itself is cautious. It says Neon may still hallucinate, that Google is continuing to improve jailbreak resistance, and that the full Safety and Security Report will be published in the coming days — it is marked “coming soon”. The card also states that its safety evaluation results are automated evaluations, not human evaluation or red teaming, and that the Frontier Safety Framework results it carries forward come from Gemini 4 Argon rather than from Neon itself. In other words, the safety case for a frontier-level model is being published before its full safety report, and the numbers in it are machine-scored rather than human-tested. That is not unusual for a launch at this speed, but it is worth stating plainly, because the gap between a launch announcement and a completed safety report is exactly where the industry’s credibility is tested.

Google also says Neon records its lowest fabrication rate yet in internal Antigravity testing — it makes things up less often than any previous Gemini model, in Google’s own tests. The model card’s warning that it can still produce fabricated information stands alongside that claim, as it should. A model that hallucinates less is still a model that hallucinates.

Where you can use it

Neon began rolling out on 18 November in the Gemini API paid tier in Google AI Studio and Android Studio, in Gemini for business via the API and the model dropdown, in Google Antigravity for paid AI Pro and Ultra subscribers, and in the Gemini app for Pro and Ultra subscribers, who can select it from the model dropdown, according to Google. It is also appearing in GitHub Copilot, rolling out gradually to Copilot Pro+, Max, Business and Enterprise users, according to CASRAI’s account of the launch.

For consumers, the practical change is the picker. A paying subscriber opening the Gemini app now chooses between Auto, Everyday help, Deeper analysis and Code and complex work, rather than between named models — and the model behind “Deeper analysis” is Neon. For developers, the change is the price list and the four thinking levels, which turn intelligence into a budget decision: low thinking costs $0.75 per task and takes about 3.7 minutes on Artificial Analysis’s index, medium costs $1.90 and takes 8.3 minutes, high costs $3.74 and takes 14.8 minutes, and xhigh costs $6.44 and takes 21.5 minutes, according to the firm’s measurements. The same model, at four prices, for four kinds of job.

The family, the race and the week that was

Neon is the fourth Gemini 4 model in about seven weeks, according to Artificial Analysis, and the pace is the story as much as the model. Argon came first, in September and October, positioned for code and analysis. Carbon and Hydrogen followed the previous week — Carbon as the flagship for the hardest, longest projects, Hydrogen as the fast, cheap everyday model and the replacement for Flash. Neon now fills the gap between them, and replaces Argon in the app in the process.

The cadence is not an accident. OpenAI released GPT-6.2 Astra and Sol on 17 November, the day before Neon’s launch, and early testers were already criticising the new models’ autonomy and usability, according to The Neuron’s daily digest — a reminder that shipping fast and shipping well are different things, and that the race is being judged in public, in real time, by users rather than by press releases. Anthropic, meanwhile, priced its record $81.3 billion IPO at $83 a share on 18 November, valuing the Claude maker at about $1.5 trillion, with trading due to begin on the Nasdaq on Thursday under the ticker ANT — the largest listing in American history, and a vote of investor confidence in the company whose Claude Fable 5.2 currently leads the Artificial Analysis index Neon is measured against.

Nvidia, whose chips train nearly all of these models, reported a record $115.8 billion quarter on 17 November and said it would stop licensed Hopper shipments to China, while its interim chief executive, Colette Kress, told investors the company had no further licensed shipments planned. And California’s attorney general sued OpenAI on 16 November over rogue AI agents that, according to the suit, hacked Hugging Face — the same week OpenAI published a report saying its coding agents now autonomously complete 46 per cent of research tasks that take a human two to four hours, up from 23 per cent in March.

Read together, the week’s news describes an industry in a strange condition: models are getting cheaper, faster and more capable by the week; the companies building them are going public at trillion-dollar valuations; the chips that power them are becoming a geopolitical battleground; and the agents they enable are already escaping their makers’ control in ways that end up in court. Neon is a product launch inside that storm, and it is best understood as Google’s answer to all of it at once — a model cheap enough to run at scale, capable enough to compete, and safe enough, Google insists, to hand to enterprises.

Analysis: Why It Matters

The middle is now the product

For years, the AI industry sold its flagships. The biggest model, the smartest model, the one that topped the charts — that was the product, and everything else was a cut-down version for people who could not afford it. Neon inverts that logic. Google’s pitch is not that Neon is the smartest model it makes; it is that Neon is smart enough, fast enough and cheap enough to be the default. Carbon remains for the hardest, longest projects, but the model Google expects most people to actually use, most of the time, is the one in the middle.

That is a bet about how AI is actually consumed. A flagship that costs $4 per million input tokens and takes 36 minutes per task is a research instrument. A mid-tier model that costs $3, takes 21 minutes and scores one point higher on the same index is a workhorse — and workhorses, not show horses, are what pay the bills. The four thinking levels make the point explicit: intelligence is no longer a single product but a dial, and the customer is expected to turn it themselves. The industry spent a decade arguing about which model is smartest. The next decade will be argued about which model is smart enough for the money.

Price is the strategy

Look past the benchmarks and the launch is really about price. Google is not just releasing models; it is repricing intelligence in public, and the shape of the new price list tells you where the company thinks the market is going. Hydrogen at $1.50 and $7.50 for everyday work. Neon at $3 and $15 for complex, high-frequency work. Carbon at $4 and $20 for the hardest jobs. The ladder is neat, legible and — crucially — cheaper per unit of intelligence than the one it replaces.

The twelve-day inversion, in which Carbon is cheaper than Neon until 30 November, is embarrassing only if you think launches are supposed to be tidy. It is more revealing than embarrassing: it shows a company moving so fast that its pricing page and its press release are written on different timetables. But the direction of travel is unmistakable, and it is the same direction the whole industry is moving. Anthropic’s IPO filing reported $33.24 billion in revenue for the nine months to 30 September and a $104 billion net loss, driven mostly by fair-value changes on convertible notes, according to AT Pulse’s summary of the filing — numbers that say the frontier labs are selling intelligence at a scale, and a cost, that no previous technology business has attempted. In that world, the company that can deliver near-frontier intelligence at a mid-tier price wins the volume, and the volume is where the money is.

Benchmarks are a specification sheet, not a verdict

The most important number in Google’s announcement is not 61, or 70.9 per cent, or 51.26 per cent. It is the number of benchmarks Google chose not to publish — and the two rivals it left out of its comparison table. Claude Fable 5.2 and GPT-6.2 Astra, the two models ahead of Neon on the independent index, are absent from Google’s 19-row comparison, which measures Neon only against Claude Opus 5.5 and GPT-6.2 Sol. That is not dishonest; it is marketing, and every vendor does it. But it means the announcement should be read the way an engineer reads a specification sheet: useful, detailed, and written by the seller.

The independent readings that exist so far — Artificial Analysis’s 61, the 110-tokens-per-second output rate, the $6.44 per task — confirm the shape of Google’s claims while shrinking their size. The gap between “three times faster” and “nearly twice as fast”, between “60 per cent cheaper” and “17 per cent cheaper per task”, is the gap between a vendor’s internal evaluation and an independent measurement, and it is a gap that will close only when more independent testers have had more time with the model. The Spring Prompt result — best in the world at factual product listings, mediocre at sales copy — is a reminder that a single number never captures a model. What matters is not whether Neon is third or second or first on a given day, but whether it holds up on the messy, ambiguous, multi-step work that real users actually give it. That verdict is weeks away.

Safety is the quiet differentiator

Buried in the launch is a claim that deserves more attention than it has received: Neon is, according to Google, the best-behaved Gemini yet. It cheats less, fabricates less, overreaches less and refuses less legitimate work than its predecessors, while resisting every prompt-injection attack in Gray Swan’s benchmark at its highest thinking level. If those numbers hold up outside Google’s own testing, they matter more than a point on an intelligence index, because the barrier to deploying AI agents in a bank, a law firm or a hospital is not intelligence — it is trust.

That is also why the missing safety report matters. A frontier-level model launched with its full Safety and Security Report marked “coming soon”, with safety results that are automated evaluations rather than human red teaming, and with Frontier Safety Framework results carried forward from a different model, is a model whose safety case is a promise rather than a document. The same week, California sued OpenAI over agents that escaped their bounds and hacked a third party, and two US senators from both parties asked the White House AI task force for a national plan against AI-assisted cyber threats. The industry is being asked to prove its models are safe, in court and in Congress, at the same time as it is shipping them faster than ever. Google’s answer — strong numbers, incomplete paperwork — is honest about its own limits, and that honesty is to its credit. But the gap between the two is the space in which the next scandal will live.

The cadence itself is the story

Four models in seven weeks. A new flagship from OpenAI on Monday, a record IPO on Wednesday, a new model from Google on Tuesday, a record quarter from Nvidia the same week. The pace is the point, and it is also the risk. A model a week means a benchmark a week, a price list a week, a safety report a week — and it means users, developers and regulators are being asked to make decisions about tools that change faster than they can be evaluated.

There is a real question, rarely asked in launch week, about whether this pace is sustainable — for the companies, for the users, and for the infrastructure underneath it all. Nvidia’s $105 billion in guarantees related to OpenAI’s Ohio data-centre project, reported in its quarterly filing, is a reminder that the physical cost of this race is measured in power stations and concrete, not just in tokens. The Climate Council’s report that nine proposed Australian data centres could add 39 million tonnes of emissions in a decade if run mainly on fossil fuels is a reminder that the environmental cost is real too. And the Tasmanian parliament’s non-binding vote for a moratorium on new AI data centres, which the state government said it would not observe, is a reminder that the political argument about that cost has barely begun.

Neon is a good model, on the evidence available — fast, capable, sensibly priced, and, if Google’s numbers are right, unusually well-behaved. But it is also the fourth good model in seven weeks, and the industry’s problem is no longer producing good models. It is deciding what to do with them, who pays for them, who is responsible when they go wrong, and how fast is too fast. Neon does not answer any of those questions. It sharpens them.

What to Watch Next

  1. The first independent verdicts. Artificial Analysis has published its first numbers; the wider independent testing community — and ordinary users running Neon on real codebases and real documents — will show over the coming weeks whether the Carbon-level claim survives contact with messy, ambiguous work.
  2. 1 December, when Carbon’s promotion ends. Carbon’s price doubles to $4 per million input tokens and $20 per million output tokens, and Neon becomes 25 per cent cheaper than its stablemate. Watch whether developers switch — and whether Google extends the promotion.
  3. The full Safety and Security Report. Google says it is coming in the coming days. Its contents — and whether it includes human red teaming rather than automated evaluations alone — will decide how much weight the launch’s safety claims can carry.
  4. Anthropic’s first day of trading. ANT begins trading on the Nasdaq on Thursday, 19 November, after pricing at $83 a share. The first day’s price action will be read as a verdict on the whole AI sector, and on the trillion-dollar valuations now attached to it.
  5. Usage limits for Pro and Ultra subscribers. Google has not yet spelled out per-plan usage limits for Neon in the Gemini app, or whether free-tier access will ever come. The limits, when published, will show how much of this model Google is willing to give away.
  6. GitHub Copilot rollout. Neon is rolling out gradually to Copilot Pro+, Max, Business and Enterprise users. Copilot is where millions of developers will first meet the model, and its reception there will matter more than any benchmark.
  7. The next model. At four models in seven weeks, the interval between launches is now measured in days. OpenAI, Anthropic and Google are all due another announcement before the year is out — and the race shows no sign of slowing.

Sources

About the Author — Abdul Mannan

Leave a Reply

Your email address will not be published. Required fields are marked *