Google launches Gemini 4 Neon: a cheaper, faster middle model that scores near the top. What Google claims, what independent testers found, and what it means for your business
Updated 19 Nov 2026, 12:09 SAST
The short version: Google has released Gemini 4 Neon, a new artificial intelligence (AI) model that sits between its fast, cheap Gemini 4 Hydrogen and its most powerful Gemini 4 Carbon[1]. Google says Neon is built for complex software work, office "knowledge work" and science tasks, and costs $3 per million input tokens and $15 per million output tokens through its developer service[1][2]. Artificial Analysis, an independent firm that benchmarks AI models, scores Neon at 61 on its Intelligence Index at the highest thinking setting, third overall, just behind OpenAI's GPT-6.2 Astra and Anthropic's Claude Fable 5.2[5]. It is also one point ahead of Google's own flagship Carbon, at a lower cost per task[5]. Paying Google AI Pro and Ultra subscribers can now pick Neon in the Gemini app[1]. Many of the headline benchmark wins come from Google itself, and some rival models still beat Neon on specific tests[6]. Our analysis explains how a small business should judge it.
What Google launched
Google announced Gemini 4 Neon in a blog post by Koray Kavukcuoglu, a senior vice-president at Google DeepMind and Google's chief AI architect, dated 18 November 2026[1]. The post describes Neon as "bringing frontier-level intelligence with fast end-to-end task completion" across complex software engineering, knowledge work and science workflows[1]. "Frontier" is the industry word for the most capable models on the market.
The first news reports appeared late on Wednesday evening, South African time. The Decoder, a German AI news site, published its report at 21:03 UTC on 18 November and 9to5Google, a site that covers Google, at 21:34 UTC[10][8]. We worked out the South African times by adding two hours to the publication times in each page's code, which makes them 23:03 and 23:34 SAST (South African Standard Time) on Wednesday.
Google says Neon is rolling out in four places at once[1]:
- For developers, in the Gemini application programming interface (API), the service that lets programmers build Google's models into their own software, through Google AI Studio and Android Studio.
- For companies, in Gemini for business, through the API or the model drop-down menu.
- In Google Antigravity, Google's AI coding tool, for paying Google AI Pro and Ultra subscribers.
- In the Gemini app, where Pro and Ultra subscribers can select Neon from the model menu.
How Neon fits in the Gemini 4 line-up
Google now sells several Gemini 4 models with element names. Google's own description is that "Carbon tackles your hardest, long-horizon projects and Hydrogen handles fast everyday execution", while Neon "balances depth and efficiency"[1]. Google's developer documentation says Neon sits between Hydrogen and Carbon, and advises developers to leave Hydrogen as the everyday default, and to reserve Carbon for the hardest problems[2].
In the Gemini app, 9to5Google reports that Neon replaces an earlier model called Argon, and that the older Gemini 3.1 Pro is "finally going away"[8]. According to 9to5Google, the app's model menu for paying users now reads: Auto; Everyday help (4 Hydrogen); Deeper analysis (4 Neon); and Code and complex work (4 Carbon)[8]. VentureBeat reports that Logan Kilpatrick of Google DeepMind posted on X that Pro subscribers can choose Neon in the Gemini app, while Carbon needs the more expensive Ultra plan[6]. He also admitted, VentureBeat reports, that Google was having "a little too much fun with naming"[6].
Google's model card, the technical summary it publishes for each model, says Neon is "adapted from Gemini 4 Carbon"[3]. Its knowledge cut-off, the date after which it has no built-in knowledge, is March 2026[3].
The fourth Gemini 4 model in weeks
Neon arrives fast on the heels of its siblings. The Decoder calls it Google's "fourth new Gemini 4 model in four weeks"[10], while Artificial Analysis counts it as the fourth "in just seven weeks"[5]. VentureBeat sets out the order: Argon about seven weeks ago, at first only for approved cybersecurity partners; then Carbon on 28 October, which opened to paying customers on 10 November[6]. The two counts differ because they start from different launches, so we give both.
What it costs
Google lists Neon at $3 per million input tokens and $15 per million output tokens, with a one-million-token "context window", the amount of text and other material the model can hold in view at once[1][2]. A token is a small chunk of text, roughly a short word or part of a longer one. Repeated input that Google can "cache" costs $0.15 per million tokens, which Google calls a 95% discount[1].
Three practical points from Google's developer documentation matter for anyone budgeting[2]:
- Neon is available only on the paid tier of the Gemini API. There is no free developer tier for this model.
- Google bills every token the model processes, including "thinking tokens", the hidden working-out a model does before it answers.
- Pricing is flat across the whole one-million-token window, rather than rising for long requests.
How the price compares
Neon is not the cheapest option, even within Google's range. The Decoder reports that Hydrogen costs $1.50 and $7.50 per million input and output tokens, half Neon's price[10]. Carbon and Argon are on an introductory price of $2 and $10 until 30 November, rising to $4 and $20 from 1 December, according to 9to5Google and The Decoder[8][10]. So Neon is dearer than Carbon for the rest of this month and 25% cheaper per token after that[8].
Against rivals, SiliconANGLE, a technology news site, says Neon sits "right in between" OpenAI's GPT-6.2 Sol, at $2 and $10, and Anthropic's Claude Opus 5.5, at $4 and $20[7]. VentureBeat notes that Neon's headline rates are 50% higher than those of Sol and of Anthropic's Claude Sonnet 5.5, but 25% below Opus 5.5[6]. VentureBeat also points out that OpenAI charges more for very long Sol requests: above 272,000 input tokens, Sol costs $4 and $15 per million for the whole request[6].
Headline token prices only tell part of the story. What a business actually pays depends on how many tokens a model burns to finish a job. VentureBeat calculates that, for the same mix of input and output, Neon would need to use "roughly one-third fewer tokens than Sol or Sonnet" to wipe out its 50% price premium[6]. That is something only testing on your own work can show.
How it performs, and who says so
Google's own numbers
Google's launch post is full of rankings. It says Neon ranks second on the Vals Index with 70.9% and first on two of its parts, the Legal Research Bench and the Tax Agent Bench[1]. Google says Neon is first on Zapier's AutomationBench at 56.3%, first on the Artificial Analysis Coding Agent Index at 58.6, and first on Andon Labs' Vending-Bench 2, a long simulation of running a small business, with $16,843[1]. Google also says Neon's cost per task at its default "high" thinking level is 32% lower than Carbon's[1].
These are Google's own selections and readings. SiliconANGLE stresses that the comparisons with Claude Opus 5.5 and GPT-6.2 Sol "were from Google's own tests"[7]. VentureBeat adds that the public snapshot of the Vals Index used for rival scores "predates Neon's release"[6].
What the independent testers found
Artificial Analysis published its own write-up on launch day[5]. Its main findings:
- Intelligence. Neon at its highest setting, "xhigh", scores 61 on the firm's Intelligence Index. That is one point ahead of Carbon at its high setting and third overall, behind GPT-6.2 Astra and Claude Fable 5.2 at their maximum settings[5]. SiliconANGLE reports those two leaders' scores as 62 for Astra and 65 for Fable 5.2[7].
- Cost per task. At xhigh Neon costs $6.44 per task on the firm's test set, about 17% less than Carbon's $7.76 at Carbon's normal, non-discounted price[5].
- Speed. Neon takes about 21 minutes per task against about 36 minutes for Carbon, roughly 40% less, because it writes out its answer at about 110 tokens per second against about 57 for Carbon[5].
- Where the saving comes from. Neon actually uses about 15% more output tokens per task than Carbon. Artificial Analysis says the lower cost comes "purely" from the 25% lower token prices[5].
On coding agents, Artificial Analysis says Neon, running in Google's Antigravity command-line tool, scores 59 on its Coding Agent Index, three points ahead of Carbon[5]. It puts that "in line with" Anthropic's Claude Code running Claude Fable 5.2 (58) and Claude Opus 5.5 (58)[5]. The firm measured an average cost per task of $6.86 for Neon in Antigravity, against $12.61 for Opus 5.5 and $15.47 for Fable 5.2 at their settings[5].
Where rivals still win
Neon does not top every chart. VentureBeat's reading of Google's own comparison chart shows Anthropic's Opus 5.5 narrowly ahead of Neon on Terminal-Bench 4.1, a test of work in a computer terminal, by 65.8% to 65.2%[6]. OpenAI's Sol led Neon on Chartography, a test of reading complex professional charts, by 75.1% to 73.5%[6]. On Cognition's FrontierCode 2.0 coding test, VentureBeat reports Neon at 46.2% and Opus at 47.3% in Google's comparison[6].
On Zapier's AutomationBench, which measures how well a model moves work between business apps, the gap ran the other way: VentureBeat reports 56.3% for Neon against 42.5% for Opus and 41.3% for Sol[6].
Speed claims also need care. A quote from Thai Tran, AI product lead at the work-search company Glean, in Google's announcement says Neon "matched or exceeded Gemini 4 Carbon on overall quality at 3x the speed and 60% of the cost", according to VentureBeat[6]. SiliconANGLE describes Google's claim as about three times as fast "at costs of about 60% lower"[7]. Those are not the same thing: 60% of the cost is a 40% saving, not a 60% one. We go with the wording VentureBeat quotes directly. VentureBeat also notes that Google's post gives no completion times in minutes for a matched Neon and Carbon job[6]. It cites the tracker Priors as measuring about 57 seconds before Neon produces its first token at xhigh[6].
The "harness" effect in coding
One of the more telling numbers in the launch is about the software around the model rather than the model itself. On a coding test called SWE-sweep, Google reports that Neon solved 12.8% of tasks with a basic open agent setup called mini-SWE, and 39.0% inside Google's own Antigravity tool[1]. The New Stack, a developer news site, points out that it was the same model in both runs[9].
The New Stack reports that Google's model card ran the same comparison for rivals. Claude Opus 5.5 rose from 3.6% with mini-SWE to 10.1% with Anthropic's Claude Code. OpenAI's GPT-6.2 Sol went the other way, from 11.6% to 8.3% with OpenAI's Codex[9]. The New Stack cautions that Google's comparison "does not isolate which changes account for the differences" and that Neon was not tested in other companies' tools[9].
The "harness" is the program that feeds the model files, runs its tools and checks its work. Google says Antigravity can now pair Neon as a planner with Hydrogen as a faster worker in a feature called Orchestrator[4]. Neon handles planning and harder reasoning, Hydrogen does routine edits and test runs, and the work goes back to Neon when Hydrogen gets stuck[4].
Safety and reliability claims
Google makes several safety claims, all from its own testing[1]:
- Neon recorded Google's "lowest fabrication rate yet" in internal Antigravity testing. Fabrication means making things up.
- On Gray Swan's benchmark for indirect prompt injection, where hidden instructions are planted in web pages, emails or documents the model reads, Google says Neon at its highest thinking level, with real-time safeguards switched on, resisted every attack.
- Neon keeps Carbon-level protection against cyber and biological misuse while refusing fewer legitimate requests, Google says.
The model card adds caveats. It says Neon "may exhibit some general limitations such as hallucinations", that Google is "continually working to improve jailbreak resistance", and that the full safety and security report "will be available in the coming days"[3]. Google says it did not run its full set of critical-capability safety tests on Neon. Its own assessment found that Neon's gain in general capability over Argon fell below the threshold that would require them, so it carried forward Argon's results[3].
A separate version, Gemini 4 Neon for cyber, will reach vetted security defenders in Google's Fairwind Program "in the coming days", the company says[1].
What it does in the Gemini app
For ordinary subscribers, Google pitches Neon as better at writing, reading uploads such as spreadsheets, PDFs and photos, and pulling details from Gmail and Docs[1]. Its examples are everyday business chores: "analyzing billing statements to spot why a utility charge spiked or cross-checking multiple receipts and documents to negotiate a refund"[1]. VentureBeat notes that Google offers these as examples and "gives no success rate" for catching billing errors[6].
What we could not verify
- South African availability and rand pricing. None of the sources we read say whether Neon is available to Gemini app subscribers in South Africa on day one, or what it costs in rand. Google prices the API in US dollars[1][2].
- Real-world reliability. Almost all quality figures come from benchmarks, chosen and in many cases run by Google. Independent results so far come mainly from Artificial Analysis[5].
- The full safety report. Google says it is still to come[3].
- Exact launch time. Google's post carries a date but no time. Our SAST times come from the first news reports, as explained above.
Why it matters for your business (analysis)
Analysis This section is our analysis, based on the reporting above. It is not advice to buy or switch any product.
The middle tier is where most work happens
For most small and medium businesses, the most expensive AI model is rarely the right default. Most daily tasks sit in the middle: summarising a contract, checking a spreadsheet, drafting a report, fixing a bug. Google says outright that Neon is aimed at that middle[4]. If the independent figures hold up, a business can get close to top-tier quality on routine but tricky tasks without paying top-tier prices. On Artificial Analysis's tests, the gap between first and third place on its Intelligence Index is four points[5][7].
Price per token is not price per job
The most useful lesson in the launch coverage is that headline prices can mislead. Neon uses more output tokens than Carbon but still costs less per task, because each token is cheaper[5]. Against OpenAI's Sol and Anthropic's Sonnet, Neon is 50% dearer per token, so it has to finish the job with far fewer tokens to come out ahead[6]. For a small firm, the practical step is to run the same five or ten real tasks through two or three models and compare total cost, time and how much fixing the result needs. Use your own invoices, your own spreadsheets and your own code. Remember that Google bills the model's hidden "thinking" too[2].
The tool around the model matters as much as the model
The SWE-sweep results suggest that the same model can score very differently depending on the software wrapped around it[9]. For a business, that means which model is best? is the wrong first question. A better one is which model works best inside the tools my team already uses? A firm that runs its work in Google Workspace may get more from Neon inside Gemini than a leaderboard suggests. A firm built around another vendor's tools may find the opposite.
Treat safety claims as a starting point
Google's prompt-injection results are encouraging, but they come from Google's own testing, and the full safety report is not yet out[1][3]. Any business letting an AI model read emails, web pages or supplier documents on its behalf should still keep a person checking anything that moves money, sends messages or changes records. Indirect prompt injection is exactly the kind of attack Google is testing against.
Expect the menu to keep changing
Google has shipped four Gemini 4 models in a matter of weeks and retired older ones from its app menu[8][10]. Carbon's price will double on 1 December[8]. Anyone building a process around one named model should write down which model and version they use, check prices monthly, and avoid locking staff training to a model name that may vanish.
Watch, test, do not rush
For most South African small businesses, nothing needs to change today. Paying Gemini subscribers can try Neon from the model menu if it is offered to them. Developers can test it through the paid API. Everyone else can wait for more independent results and the full safety report before moving any important work to it.
Sources
- ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l ↩m ↩n ↩o ↩p ↩q ↩r ↩s Google (Koray Kavukcuoglu, SVP, Google DeepMind), “Introducing Gemini 4 Neon”, 18 November 2026.
- ↩a ↩b ↩c ↩d ↩e ↩f Google AI Studio documentation, “What's new in Gemini 4 Neon”, accessed 19 November 2026.
- ↩a ↩b ↩c ↩d ↩e ↩f Google DeepMind, “Gemini 4 Neon – Model Card”, published November 2026, accessed 19 November 2026.
- ↩a ↩b ↩c Google Antigravity blog, “Gemini 4 Neon in Google Antigravity”, 18 November 2026.
- ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l ↩m ↩n Artificial Analysis, “Google continues to accelerate its release cadence with Gemini 4 Neon”, 18 November 2026.
- ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l ↩m ↩n ↩o ↩p ↩q VentureBeat (Carl Franzen), “Google launches Gemini 4 Neon with Opus 5.5-level coding performance at 25% lower API token prices”, 18 November 2026.
- ↩a ↩b ↩c ↩d ↩e SiliconANGLE (Mike Wheatley), “Google's latest midtier model Gemini 4 Neon is three times faster than Carbon and wins on some metrics”, 18 November 2026 (updated 21:55 EST).
- ↩a ↩b ↩c ↩d ↩e ↩f ↩g 9to5Google (Abner Li), “Google launches Gemini 4 Neon with Carbon quality at Flash speeds, live for AI Pro”, 18 November 2026.
- ↩a ↩b ↩c ↩d The New Stack (Amanda Caswell), “Google's Gemini 4 Neon tripled its coding score without changing the model”, 19 November 2026.
- ↩a ↩b ↩c ↩d ↩e The Decoder (Matthias Bastian), “Google ships Gemini 4 Neon, its fourth new Gemini model in a month”, 18 November 2026.
This article is for information only and is not financial or investment advice.
Spotted an error? Tell us and we'll correct it.
Featured deep dives
Longer follow-ups on the stories behind the week's headlines, with every source linked. Free to read.