Google Launches Gemini 4 Neon, Its Fastest Frontier AI Model Yet, at  per Million Tokens
Technology 8 min read

Google Launches Gemini 4 Neon, Its Fastest Frontier AI Model Yet, at $3 per Million Tokens

By November 22, 2026

Google has unveiled Gemini 4 Neon, the fourth model in its Gemini 4 family in seven weeks, promising top-tier reasoning at the speed of a smaller model — and at a price that undercuts its own flagship.

By The BNN Media
The Best News Network
News Beyond the Borders

Mountain View, United States — Google has launched Gemini 4 Neon, a new artificial intelligence model that the company says delivers frontier-level reasoning for complex, high-frequency work — coding, research and enterprise tasks — faster and more cheaply than its larger models, with pricing starting at $3 per million input tokens.

Google Launches Gemini 4 Neon, Its Fastest Frontier AI Model Yet, at  per Million Tokens
A Google office building with the company’s name on its glass facade. (File photo; the Google logo shown is the building’s own signage.) Photo: K927 / dehayf5mhw1h7.cloudfront.net

Neon, introduced on Tuesday 18 November, is the fourth model in the Gemini 4 family to arrive in seven weeks, following Argon in late September, Carbon in late October and Hydrogen earlier in November. Google describes it as the middle of the range: Carbon tackles the hardest, long-running projects, Hydrogen handles fast everyday execution, and Neon balances depth and efficiency for work that is both demanding and constant.

What Google launched

Neon is built for what Google calls “complex, high-frequency workflows” — long-running coding jobs, analytical reasoning and production-grade enterprise work. It carries a context window of one million tokens, meaning it can take in and reason across very large documents, codebases and datasets in a single session, and it offers four thinking levels — low, medium, high and extra high — so users can trade intelligence against cost and speed for each task.

The model began rolling out on 18 November in the Gemini API through Google AI Studio, in Gemini for business, in Google Antigravity for paid AI Pro and Ultra subscribers, and in the Gemini app for Pro and Ultra subscribers. A version built for cyber defence, Neon for cyber, is being rolled out to trusted defenders in the coming days through Google’s Fairwind Program.

Google says Neon is designed to sustain performance through long, tool-heavy tasks from start to finish — whether working alone or acting as a fast, high-judgement orchestrator directing parallel Hydrogen sub-agents inside Antigravity, its AI coding platform. It cuts down on redundant file reads, shares progress as it works and handles the compression of long conversations reliably, according to the company.

Neon is Google’s attempt to answer the question every AI lab now faces: how to make a model that is both clever and cheap enough to run all day.

How fast, how smart — Google’s own numbers

Google’s launch claims place Neon near the top of the independent leaderboards it cites. On the Artificial Analysis Intelligence Index, the company says Neon scores 61 at its extra-high thinking level, ranking third overall — just behind larger, slower models. On the Vals Index it ranks second with 70.9 per cent, and it takes first place on both the Legal Research Bench and the Tax Agent Bench.

For coding, Google says Neon ranks first overall on the Artificial Analysis Coding Agent Index with 58.6 points, scores 65.2 per cent on Terminal-Bench 4.1, and reaches 46.2 per cent on FrontierCode 2.0 at its default high setting. On long-horizon business simulation, it ranks first on Zapier’s AutomationBench (56.3 per cent) and on Andon Labs’ Vending-Bench 2, where it earned $16,843 running a simulated shop.

  • Artificial Analysis Intelligence Index: 61 at extra-high thinking, ranked #3 overall.
  • Artificial Analysis Coding Agent Index: 58.6, ranked #1 overall.
  • Terminal-Bench 4.1: 65.2 per cent; FrontierCode 2.0: 46.2 per cent at default high setting.
  • Vals Index: #2 (70.9 per cent); Legal Research Bench and Tax Agent Bench: #1.
  • Zapier AutomationBench: 56.3 per cent, ranked #1; Vending-Bench 2: $16,843, ranked #1.
  • Terminal-Bench Science: 70.5 per cent; RiemannBench: 84.0 per cent; HLE-Diamond: 71.3 per cent — all described by Google as best in class.

These are vendor-reported results, and no independent benchmarks of Neon have been published yet. They should be read as Google’s account of its own model’s performance, not as verified fact. But the pattern of the claims is consistent: Neon is pitched as a model that does the work of a flagship while behaving like a workhorse.

On Google’s own figures, Neon is the fastest model the company has yet put near the top of the leaderboards — but the numbers are Google’s, and independent testing has yet to catch up.

What it costs

Neon is priced at $3 per million input tokens — $0.15 per million for cached input, a 95 per cent discount — and $15 per million output tokens. Google says that at its default high thinking level, Neon delivers a high intelligence score at a 32 per cent lower cost per task than Carbon, and roughly 50 to 80 per cent lower cost per task than top competing frontier models.

The pricing picture is about to shift. Argon and Carbon currently cost $2 per million input tokens and $10 per million output under an introductory offer that ends on 30 November; from 1 December they rise to $4 and $20, at which point Neon becomes 25 per cent cheaper per token than both, according to 9to5Google.

Neon also replaces Argon in the Gemini app’s model picker for paid subscribers, and Google has made 3.5 Flash-Lite the automatic choice for paid users. The message is clear: Google is tidying its line-up around three models — Carbon for the hardest jobs, Neon for the demanding everyday ones, Hydrogen for speed.

A data centre server room with rows of server racks. (File photo) Photo: mop.dev

Built for code

Coding is where Google is making its strongest pitch. Neon is engineered for long-running agent workflows, and in Antigravity it can act as an orchestrator, planning work and directing Hydrogen sub-agents to carry it out in parallel. Google says it handles multi-file debugging and refactoring across whole repositories with targeted, surgical edits.

On SWE-sweep, a benchmark for real-world software engineering, Google says Neon leads on the smaller harness with 12.8 per cent and more than triples to 39.0 per cent inside Antigravity — a figure the company presents as evidence of what its own tooling unlocks. It also says Neon records the lowest rate of fabrication — inventing results — of any Gemini 4 model in internal Antigravity testing.

For developers, the appeal is practical: a model that can be left to work through a long task, check its own progress and finish the job without being babysat. Whether it does that reliably in the wild, outside Google’s own tests, is the question early users will now answer.

Cyber defence and safety

Google has also positioned Neon as a cyber-defence tool. The company says it produces more secure code than Carbon across easy, medium and hard challenges on AutoBaxBench, reaching the top score on the hard tier at 76 per cent. On WebExploitBench, a penetration-testing benchmark, it scores 51.26 per cent at half the cost of competing large models. On vulnerability finding it reaches 90.5 per cent on SEC-bench Pro, and on patching it scores 70.7 per cent on CWE-bench v2, a 3.4 per cent improvement over Carbon.

On safety, Google says Neon maintains Carbon-level protections against severe misuse and prompt injection, while declining far less legitimate work in sensitive areas. It says the model matched Carbon with 0.0 per cent cheating on Terminal-Bench 4.0, took shortcuts on just 2.45 per cent of tasks in its coding-cheating evaluation — the lowest rate of any Gemini 4 model — and achieved a 99.2 per cent pass rate on its agent honesty evaluation. On Gray Swan’s indirect prompt injection benchmark, Google says Neon resisted every attack at its highest thinking level.

These, again, are the company’s own measurements. But they show where Google thinks the argument for Neon is won: not just on raw capability, but on behaving safely and honestly while doing long, unsupervised work.

Google is selling Neon as the model you can trust to work alone — and its safety claims are as central to the pitch as its benchmark scores.

The Gemini 4 family in seven weeks

Neon’s arrival completes a rapid-fire release schedule that has seen Google ship four models in under two months:

  • Argon — 30 September: the first Gemini 4 model, aimed at complex, long-running tasks, with output of up to one million tokens.
  • Carbon — 28 October: described by Google as its strongest model, with an emphasis on cybersecurity and long-running tasks; widely available since 10 November.
  • Hydrogen — 12 November: a fast, cheap model for everyday programming work and AI agents.
  • Neon — 18 November: the middle model, combining Carbon-level intelligence with Flash-like speed and a lower price.

The pace reflects the intensity of the AI race. Google’s rivals have been shipping at a similar tempo — OpenAI launched its GPT-6.2 Astra and Sol models on the same week — and each release is now judged as much on cost per task as on capability.

A launch into a falling market

Neon arrived in a week when technology shares suffered their worst week since 2020, with investors questioning the valuations built on the AI boom. Against that backdrop, a model pitched on efficiency rather than raw power is a deliberate choice: Google is arguing that the next phase of the AI race will be won on cost, reliability and usefulness, not on ever-larger models.

For now, access is limited. In the Gemini app, Neon is available only to paying AI Pro and Ultra subscribers; developers need the paid tier of the Gemini API. Whether it earns a wider audience will depend on what those early users find — and on whether independent benchmarks back up Google’s claims.

Neon is Google’s bet that the AI race is entering its practical phase — and that the model that wins will be the one that does serious work, quickly, at a price businesses can afford.

A software developer working on code at a computer monitor. (File photo) Photo: TalentCrowd

By The BNN Media
The Best News Network
News Beyond the Borders

Leave a Reply

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