Skip to content
, ,

Gemini 4 Neon Is Here: Price, Benchmarks and How to Use It

10–15 minutes
Woman at a neon-lit desk holding a card with the Gemini logo, hook Gemini 4 Neon beats Opus?

Google released Gemini 4 Neon on 18 November 2026. It is the fourth Gemini 4 model in about a month and sits between the fast Hydrogen and the top-end Carbon. Google’s pitch is simple: close to Carbon quality, much faster, and cheaper per task.

It is live today in the Gemini app for AI Pro and Ultra subscribers, in Google Antigravity, and in the paid Gemini API at $3 per million input tokens and $15 per million output tokens.

On Google’s charts it beats Claude Opus 5.5 on several coding tests and ranks first on the Artificial Analysis Coding Agent Index.

This guide covers what Neon is, where you can use it, the full price picture, every benchmark Google published, how to call it from code, and when to pick it over Hydrogen, Carbon or Claude.

TL;DR

  • What: Gemini 4 Neon, a new mid-to-top Gemini 4 model announced 18 Nov 2026 by Google DeepMind’s Koray Kavukcuoglu.
  • Price: $3 input / $15 output per 1M tokens, $0.15 cached input. That is 25% below Carbon’s price from 1 December ($4 / $20).
  • Context: 1M tokens; reads text, images, PDFs, spreadsheets, audio and video.
  • Where: Gemini app (“Deeper analysis” in the model picker) for AI Pro and Ultra, Antigravity for Pro and Ultra, Gemini API paid tier in AI Studio, Android Studio, Gemini for business.
  • Model ID: gemini-4-neon.
  • Benchmarks (Google’s own): #1 Artificial Analysis Coding Agent Index (58.6), Terminal-Bench 4.1 65.2%, CursorBench 4.1 65.4% vs Opus 5.5’s 58.1%.
  • Free? Not on the free Gemini plan or free API tier at launch.

What Is Gemini 4 Neon?

Neon is Google’s new “balanced” model in the Gemini 4 family. Google describes it as bringing frontier-level intelligence with fast end-to-end task completion to software engineering, knowledge work and science.

Here is how Google now splits the family:

Model Role API price per 1M (input / output)
Gemini 4 Hydrogen Fast everyday work, replaces Flash $1.50 / $7.50
Gemini 4 Neon Depth plus speed for complex, frequent workflows $3 / $15
Gemini 4 Argon Code and analysis $2 / $10 until 30 Nov, then $4 / $20
Gemini 4 Carbon Hardest long-horizon projects $2 / $10 until 30 Nov, then $4 / $20

Prices: Google’s Neon post, plus 9to5Google and The Decoder for Hydrogen, Argon and Carbon. See my Gemini 4 Hydrogen guide and Gemini 4 Carbon guide for those models.

A few things stand out about Neon:

  • Speed: Google and 9to5Google describe it as “Carbon quality at Flash speeds”. Google gave no tokens-per-second figure. One early tester measured about 127 tokens per second through OpenRouter.
  • Four thinking levels: high is the default and xhigh is the maximum. Lower levels cut cost and wait time.
  • Cheaper per task: Google says Neon at high costs 32% less per task than Carbon, and 50% to 80% less than top rival frontier models on the Vals and Agent Arena tests.
  • Better agent habits: fewer repeated file reads, progress updates while working, context compaction in long sessions, and the lowest fabrication rate Google has recorded in Antigravity testing.
  • Multimodal: spreadsheets, PDFs, charts, audio and video in one 1M-token context, with structured outputs like reports and tables.

Where You Can Use It

Surface Who gets it Notes
Gemini app (web, Android, iOS) Google AI Pro and AI Ultra Pick “Deeper analysis” in the model menu. Also powers Tasks.
Google Antigravity Paid AI Pro and Ultra subscribers Can direct parallel Hydrogen subagents
Gemini API / Google AI Studio Paid tier developers; Pro and Ultra subscribers can try it in AI Studio Model ID gemini-4-neon
Android Studio Paid tier developers Agent mode model choice
Gemini for business Workspace business customers API and model dropdown
OpenRouter Anyone with credits Listed as google/gemini-4-neon
Neon for cyber Fairwind Program members Rolling out “over the next few days”; request via your Cloud account team

Google did not mention free Gemini users, so assume Neon is paid-only for now. Free users still get Hydrogen through Auto. If you are deciding whether to upgrade, my Google AI Pro vs Ultra comparison covers what each plan includes.

Model picker change: 9to5Google reports the Gemini app menu now reads Auto, Everyday help (Hydrogen), Deeper analysis (Neon) and Code & complex work (Carbon). It also says Gemini 3.1 Pro is being retired from the app.

Price and What a Task Costs

Item (per 1M tokens) Gemini 4 Neon
Input $3.00
Cached input $0.15 (95% off)
Output (thinking tokens included) $15.00
Context window 1,000,000 tokens
Free API tier Not at launch (paid tier only)

Source: Google’s announcement and TestingCatalog, 18 November. Google’s main API pricing page had not added any Gemini 4 rows when I checked on 19 November, so batch and long-context prices are not confirmed yet.

Example costs

  • A 20K-token prompt with a 4K-token answer: $0.06 + $0.06 = $0.12.
  • A coding agent session reading 2M tokens (1.5M cached) and writing 300K: $0.225 cached + $1.50 new input + $4.50 output = about $6.20.
  • Arena’s measured figure: a median agent task cost $2.82 on Neon, which Arena says is 53% less than Claude Opus 5.5.

Tip: thinking tokens are billed as output. Dropping from high to a lower thinking level is the biggest cost lever. My reasoning effort and API cost guide shows how this works across models.

Benchmarks

Most numbers below come from Google’s post. Some, like the Artificial Analysis, Vals and Arena results, are from third-party leaderboards that Google quoted on launch day.

Benchmark Gemini 4 Neon Context
Artificial Analysis Coding Agent Index 58.6 (#1) Top of the coding agent leaderboard
Artificial Analysis Intelligence Index 61 (#3, xhigh) Behind larger, slower models
Vals Index 70.9% (#2) 0.4 points behind Claude Fable 5.2 at under half the cost per task
CursorBench 4.1 65.4% Claude Opus 5.5: 58.1%
Terminal-Bench 4.1 65.2% Working in a real terminal
Arena Agent Arena 1,336 (#2) About 4,200 sessions; ahead of Carbon (1,303)
Arena Text Arena 1,518 (#3) Google holds the top three spots
Agents’ Last Exam v1.1 75.8% On par with leading models
Zapier AutomationBench 56.3% (#1) Workflow automation
Vending-Bench 2 $16,843 (#1) Running a simulated business
Terminal-Bench Science 70.5% Best in class
RiemannBench 84.0% Maths
HLE-Diamond 71.3% Hard expert questions
FrontierCode 2.0 46.2% At default high thinking
SWE-sweep 39.0% (Antigravity harness), 12.8% (mini-SWE) Long software tasks
Chartography 73.5% Reading charts

Vals also lists Neon as #1 on Legal Research Bench and Tax Agent Bench. Surge AI ranks it first on most of its finance, accounting and spreadsheet tests.

The honest summary: Neon is not Google’s smartest model, but it is the best value Google has shipped for agents and coding. The big claim is not the raw score but the cost per finished task. Artificial Analysis puts it at about 17% cheaper per task than Carbon, with slightly higher intelligence.

How to Use Neon in the Gemini App

  1. Check your plan. You need Google AI Pro or AI Ultra on the account.
  2. Open Gemini at gemini.google.com or in the Android or iOS app.
  3. Tap the model menu at the top (or beside the prompt box) and choose Deeper analysis. That is Neon.
  4. Upload files if you need to. Neon handles PDFs, spreadsheets, charts, audio and video in one chat.
  5. Use Tasks for longer jobs. Google says Tasks can now run on Neon, which stretches your usage limits further than Carbon would.

Neon can also pull from Gmail and Docs when you connect them, which helps for writing and research on your own material.

How to Use Neon in the Gemini API

  1. Get a key in Google AI Studio and turn on billing. Neon is paid-tier only.
  2. Install the SDK: pip install -U google-genai.
  3. Set the key as GEMINI_API_KEY.
  4. Call gemini-4-neon as shown below.

Python

from google import genai
from google.genai import types

client = genai.Client()  # reads GEMINI_API_KEY

resp = client.models.generate_content(
    model="gemini-4-neon",
    contents="Review this function for bugs and suggest a fix: ...",
    config=types.GenerateContentConfig(
        thinking_config=types.ThinkingConfig(thinking_level="high"),
    ),
)
print(resp.text)
print(resp.usage_metadata)  # thinking tokens are billed as output

curl

curl "https://generativelanguage.googleapis.com/v1beta/models/gemini-4-neon:generateContent" \
  -H "x-goog-api-key: $GEMINI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"contents":[{"parts":[{"text":"Explain MoE models in two lines."}]}]}'

Neon supports Google Search and Maps grounding, code execution, file search and Computer Use, according to The Decoder’s reading of the AI Studio changelog. It does not generate images or audio.

Coming from Hydrogen? On Hydrogen, Google made high the default thinking level and ignored temperature settings without any error. Test Neon the same way before you swap it into production. My Gemini API Python tutorial and function calling tutorial cover the rest of the setup.

Cursor, Cline and OpenRouter

  • OpenRouter: use model google/gemini-4-neon with any OpenAI-compatible client.
  • Cline, Roo, OpenCode: choose the Google Gemini provider, paste your AI Studio key, and type gemini-4-neon if it is not in the list yet.
  • Context setting: many tools default lower than 1M; raise it if you load big repos.

Neon in Antigravity

In Google Antigravity, Neon is available to paid AI Pro and Ultra subscribers. Google highlights two things:

  • Orchestrator mode: Neon can plan a job and hand pieces to several Hydrogen subagents working in parallel. This is cheaper than running everything on one big model.
  • Long sessions: it compacts context on its own and shares progress while working, so long refactors are easier to follow.

Google’s 39.0% SWE-sweep score was measured inside Antigravity’s own harness. For how parallel agents work there, see my Antigravity swarm and remote agents guide.

Neon vs Hydrogen, Carbon, Claude and GPT-6

Model Input / output per 1M Best for
Gemini 4 Hydrogen $1.50 / $7.50 Cheap, fast everyday tasks
Gemini 4 Neon $3 / $15 Coding agents, research, documents
Gemini 4 Carbon $2 / $10 now, $4 / $20 from 1 Dec The hardest long projects
Claude Opus 5.5 $4 / $20 Reliable coding and terminal work
Claude Fable 5.2 $8 / $40 Top-end reasoning
GPT-6 Astra $10 / $50 OpenAI’s flagship
Qwen4-Max $1.50 / $4.50 Lowest-cost frontier-class coding

Rival prices are list prices as reported in mid-November. For a wider table see my AI API pricing comparison and Qwen4-Max guide.

Who should pick which

  • Pick Neon for coding agents, spreadsheet and document analysis, and research where you want near-top quality without top-tier bills.
  • Stay on Carbon until 30 November if you already use it: its promo price is actually cheaper than Neon until then. From 1 December, Neon is the cheaper choice.
  • Pick Hydrogen for chat, quick edits and high-volume simple calls.
  • Stay with Claude Opus 5.5 if your agents are tuned for it and you rely on its terminal habits; test Neon side by side first.

Safety and the Cyber Version

Google says Neon keeps Carbon-level protection against severe misuse and prompt injection. Its published numbers include:

  • 0.0% attack success on Gray Swan’s indirect prompt injection test at the top thinking level with live safeguards.
  • 0.09% misaligned actions on Google’s agentic safety suite and 99.2% on its agent honesty test.
  • Shortcuts on 2.45% of tasks in a coding cheating test, the lowest of any Gemini 4 model.
  • Overreach cut by 15% to 35% and “under-service” (doing too little) by 43% to 56%, depending on effort level.

A separate Neon for cyber version is rolling out to Fairwind Program members for defensive security work. Google lists 76% on AutoBaxBench’s hard tier and 90.5% on SEC-bench Pro. It is not open to the public.

Limits and Unknowns

  • No free access in the Gemini app or API at launch.
  • Max output tokens, batch prices and rate limits were not in Google’s post. Check AI Studio’s model page before planning big jobs.
  • Vendor numbers: most benchmarks are Google’s own or picked by Google. Independent tests will take a few days.
  • No local version. Neon is closed and runs only on Google’s servers.
  • Usage limits in the Gemini app for Pro and Ultra were not published.

Troubleshooting

Problem Likely cause Fix
404 or “model not found” Typo, old SDK, or free-tier project Use exactly gemini-4-neon, update google-genai, turn on billing
429 RESOURCE_EXHAUSTED Rate limit on a new paid project Back off and retry; see my rate limits guide
No “Deeper analysis” option in Gemini Free plan, or rollout not reached your account Check you are on AI Pro or Ultra, sign out and back in, update the app
Bills higher than expected Default high thinking uses many output tokens Set a lower thinking_level and check usage_metadata
Answer ignores your temperature setting Gemini 4 models may ignore sampling settings Control style with the prompt instead
Other API errors Key, region or request format See my Gemini API errors fix guide

My Take

Neon is the Gemini 4 model most people will actually want. It is quick, it scores near the top on coding and agent tests, and from 1 December it costs 25% less than Carbon.

The CursorBench lead over Claude Opus 5.5 (65.4% vs 58.1%) and the #1 coding agent index spot are the numbers to watch. If independent tests confirm them, Neon puts real price pressure on Claude for coding agents.

The catch is access: no free tier and no free Gemini app access yet. If you already pay for AI Pro, switch to “Deeper analysis” for your next big document or coding task and see how it feels.

FAQ

When was Gemini 4 Neon released?

Google announced and released it on 18 November 2026.

Is Gemini 4 Neon free?

No. It needs Google AI Pro or Ultra in the Gemini app and Antigravity, or the paid tier of the Gemini API.

How much does the Gemini 4 Neon API cost?

$3 per million input tokens, $0.15 per million cached input tokens and $15 per million output tokens. Thinking tokens count as output.

What is the Gemini 4 Neon model ID?

Use gemini-4-neon in the Gemini API, or google/gemini-4-neon on OpenRouter.

Is Neon better than Gemini 4 Carbon?

Google says Neon matches Carbon on most coding, reasoning and agent tasks while being faster and about 32% cheaper per task. Carbon is still meant for the hardest long projects.

Is Gemini 4 Neon better than Claude Opus 5.5?

On Google’s charts it scores higher on CursorBench 4.1 (65.4% vs 58.1%) and Arena says it costs 53% less per agent task. Independent tests are still coming.

How big is the context window?

1 million tokens, covering text, images, PDFs, spreadsheets, audio and video.

Can I run Gemini 4 Neon locally?

No. There are no downloadable weights. For local models, look at open models such as Qwen3.8 or Kimi K3.1.

Sources

  • Google blog: “Introducing Gemini 4 Neon”, Koray Kavukcuoglu, 18 November 2026
  • 9to5Google: “Google launches Gemini 4 Neon with Carbon quality at Flash speeds, live for AI Pro”, 18 November 2026
  • TestingCatalog: “Google launches Gemini 4 Neon at $3 per million tokens”, 18 November 2026
  • The Decoder: “Google ships Gemini 4 Neon, its fourth new Gemini model in a month”, 18 November 2026
  • micjm.com: “Gemini 4 Neon looks like an Opus 5.5 competitor” (OpenRouter speed test), 18 November 2026
  • Google AI for Developers: Gemini API pricing page, checked 19 November 2026
Free weekly email

Get new AI guides every week

One short weekly email with our newest local AI, ComfyUI and AI tool guides. No spam, no pop-ups, just the good stuff.

Free. Unsubscribe anytime in one click. We only use your email to send this newsletter. See our privacy policy.

Woman at a neon-lit desk holding a card with the Gemini logo, hook Gemini 4 Neon beats Opus?