On November 18, 2026, Google announced Gemini 4 Neon, the fourth Gemini 4 model to ship in a single month. Neon sits between Hydrogen, the fast member of the family, and Carbon, the largest. Google's pitch is simple: quality close to Carbon, ahead of it on some tests, delivered much faster and for less money.

This guide covers Gemini 4 Neon in detail: what it is, what it can do, how it scores, what it costs and where you can use it. We then run two comparisons. First against Google's own models (Hydrogen, Carbon, Argon, 3.5 Flash-Lite), then against other brands (Anthropic Claude, OpenAI GPT, xAI Grok, DeepSeek, Meta, Moonshot Kimi, Qwen). We close with a practical guide to which model fits which business task.

Note: Most benchmark results in this article were published by Google itself. Independent measurements (such as Artificial Analysis and Vals) are labelled. Prices are in US dollars per 1 million tokens, as of November 22, 2026.

Gemini 4 Neon at a Glance

Feature Gemini 4 Neon
Announced November 18, 2026
Developer Google DeepMind
API model ID gemini-4-neon
Context window 1,048,576 tokens (shared by input, thinking and output)
Max output per response 65,536 tokens (up to 1M with long decode)
Inputs Text, images, video, audio, PDF
Output Text only (no image or audio generation)
Thinking levels low, medium, high (default), xhigh
API price $3 input, $15 output, $0.15 cached input
Free tier None (paid API projects only)
Gemini app Google AI Pro and AI Ultra subscribers

Google describes Neon as "frontier-level intelligence with fast end-to-end task completion" for complex software engineering, knowledge work and science workflows. According to the model card, Neon is adapted from Carbon and shares its architecture and training data.

The Gemini 4 Family and Where Neon Fits

With Gemini 4, Google dropped names like "Pro" and "Flash" in favour of element names. Four models in a few weeks make the line-up confusing, something Google's Logan Kilpatrick has acknowledged himself.

Model Released Role Name in the Gemini app
Gemini 4 Hydrogen Nov 12, 2026 Fast everyday work; replaces "Flash" Everyday help
Gemini 4 Neon Nov 18, 2026 Mid-tier; deep analysis, code, knowledge work Deeper analysis
Gemini 4 Carbon Nov 10, 2026 (GA) Largest model; hardest long-horizon work Code & complex work
Gemini 4 Argon October 2026 Code and analysis; first offered to cyber partners Replaced by Neon

In the app, Neon replaces Argon for AI Pro subscribers, and the old Gemini 3.1 Pro is gone from the picker. For paid users, Gemini 3.5 Flash-Lite now appears as "Auto".

What It Can Do

1 million token context. Roughly 700,000-750,000 words of text, a large contract bundle, a big code repository or hours of video in one request. Pricing is flat across the whole window, with no long-context surcharge.

Multimodal input. It reads text, images, video, audio and PDF. On Google Cloud it also reads Word (.docx), Excel (.xlsx) and .odt files directly.

Tool use. Computer use, code execution, file search, function calling, structured output, URL context, and grounding with Google Search and Google Maps are all supported, which makes Neon a good fit for agent-style applications.

Four thinking levels. low and medium for quick answers, high as the default, xhigh for the hardest work. Google recommends medium for interactive coding, high for most complex tasks, and xhigh only when high falls short. Thinking tokens are billed at the output rate, so higher levels cost more. In the Gemini app, if you pick a model manually without setting a level, it defaults to low.

What is missing? No image generation, no audio generation and no Live API (real-time voice). Explicit context caching is not supported; only implicit caching is.

Benchmarks

The table below is compiled from Google's published comparison against Anthropic's Claude Opus 5.5 and OpenAI's GPT-6.2 Sol. The best score in each row is in bold. "-" means Google did not report that competitor.

Benchmark Area Gemini 4 Neon Claude Opus 5.5 GPT-6.2 Sol
Artificial Analysis Intelligence Index General 61 59 57
Vals Index v2.2 General 70.9% 70.1% 66.1%
AutomationBench (Zapier) Knowledge work, automation 56.3% 42.5% 41.3%
Agents' Last Exam v1.1 Knowledge work 75.8% 71.6% 73.0%
GraphWalks (256K-1M) Long context 95.2% 66.8% 80.1%
CursorBench 4.1 Agentic coding 65.4% 58.1% -
Terminal-Bench 4.1 Agentic coding 65.2% 65.8% 61.5%
FrontierCode 2.0 Agentic coding 46.2% 47.3% 43.8%
SWE-sweep (Antigravity) Agentic coding 39.0% 10.1% 8.3%
SWE-sweep (shared Mini-SWE) Agentic coding 12.8% 3.6% 11.6%
RiemannBench Math 84.0% 80.0% 80.0%
Terminal-Bench-Science Science 70.5% 63.3% 68.0%
HLE-Diamond Reasoning 71.3% 67.9% 64.0%
MKA Multimodal 56.5% 53.0% 52.7%
Chartography Chart reading 73.5% 66.3% 75.1%
LVBench Video 89.0% 83.7% 87.6%
CWE-Bench v2 Cybersecurity 70.7% 67.1% 66.0%
SEC-bench Pro Cybersecurity 90.5% - 88.3%
WebExploitBench v1.2 Cybersecurity 51.3% 37.1% -

What the table shows:

  • Neon leads on 16 of 19 tests. Some outlets say "17 of 19", but the table shows three losses: Claude Opus 5.5 wins Terminal-Bench 4.1 and FrontierCode 2.0, and GPT-6.2 Sol wins Chartography.
  • The biggest gaps are in long context and automation. 95.2% vs 66.8% on GraphWalks and 56.3% vs 42.5% on AutomationBench are large margins.
  • Coding is close. Opus 5.5 still edges Neon on the terminal and FrontierCode tests. The big SWE-sweep lead depends on the harness: 39% in Google's own Antigravity, 12.8% in the shared Mini-SWE harness, where the gap to GPT-6.2 Sol is only 1.2 points.
  • Read the cyber numbers carefully. Some sources attribute the SEC-bench Pro (90.5%) and CWE-Bench (70.7%) results to a specialised cyber version of Neon that is only offered to Fairwind Program members.

Other rankings Google reported: #1 on the Artificial Analysis Coding Agent Index (58.6), #2 for steerability on Agent Arena (score 1336), #1 on Vals' Legal Research and Tax Agent benchmarks, #1 on Vending-Bench 2 ($16,843), and #1 on most of Surge AI's finance, accounting and spreadsheet modelling tests.

Speed

  • According to an independent tracker, Neon produces about 110 tokens per second at xhigh. Carbon runs at 59 tokens per second at high, so Neon is roughly 1.9x faster.
  • Hydrogen is the fastest in the family at 167-171 tokens per second.
  • The trade-off: at xhigh, time to first token is about 57 seconds. That level is meant for long, hard jobs, not live chat.
  • A customer quote shared at launch (Glean) says Neon matched or beat Carbon's quality in early tests at 3x the speed and 60% of the cost.

API Pricing

Tier Input Output Cached input
Standard $3.00 $15.00 $0.15
Batch $1.50 $7.50 $0.075
Flex $1.50 $7.50 $0.075
Priority $5.40 $27.00 $0.27
  • Cached input is 95% off. Apps that resend the same system prompt or the same documents save a lot.
  • Batch and Flex are half price for non-urgent work such as overnight reports and bulk text processing.
  • Google Search grounding: 5,000 free searches per month, then $14 per 1,000.
  • Google says cost per task at high is 32% lower than Carbon. Artificial Analysis says the advantage narrows to about 17% when Neon runs at xhigh.

Watch the pricing timing: Carbon and Argon are on introductory pricing ($2 / $10) until November 30, 2026. Until then, Neon looks 50% more expensive per token than Carbon. From December 1, Carbon rises to $4 / $20 and Neon becomes 25% cheaper than Carbon.

Compared With Google's Own Models

Hydrogen Neon Carbon Argon 3.5 Flash-Lite
Role Fast, everyday Mid-tier, deep analysis Largest, hardest work Code and analysis Fastest, cheapest
Input / output price $1.50 / $7.50 $3 / $15 $2 / $10 (to Nov 30), then $4 / $20 Same as Carbon $0.30 / $2.50
Free API tier Yes No No No Yes
Thinking levels low-xhigh (4) low-xhigh (4) low-high (3) low-high (3) minimal-high
Context 1M 1M 1M 1M -
AA Intelligence Index 54 61 60 - -
Vals Index 68.5% 70.9% 70.5% - -
Speed (tokens/s) 167-171 ~110 ~59 - Fastest
Gemini app Including free AI Pro, Ultra AI Ultra only Removed Auto

Neon or Carbon? On independent indexes Neon edges Carbon (61 vs 60) and is about twice as fast. Google has not published a head-to-head table for the two, so it is unclear where Carbon really pulls ahead. In a small DataCamp test (building a chess game page), Neon produced a working page in 3 of 3 runs, Carbon in 1 of 3, because Carbon hit the single-response output cap. For most work today, Neon looks like the better choice. Carbon remains the option for very long, hardest-tier projects, and it is cheap until November 30 thanks to the promo price.

Neon or Hydrogen? Hydrogen is half the price, faster, and has a free API tier. Use Hydrogen for high-volume work like customer chat, classification and short summaries; use Neon for analysis, reports, code and multi-step agent work. They share the same request schema, context and tools, so they can run in the same app. Antigravity's Orchestrator mode already does this: Neon plans, Hydrogen executes.

Compared With Other Brands

Price Comparison

Model Company Input Output
Gemini 4 Neon Google $3.00 $15.00
Claude Opus 5.5 Anthropic $4.00 $20.00
Claude Sonnet 5.5 Anthropic $2.00 $10.00
Claude Haiku 5.5 (up to 100K) Anthropic $0.10 $0.50
Claude Fable 5.2 Anthropic $8.00 $40.00
GPT-6.2 Sol (up to 272K) OpenAI $2.00 $10.00
GPT-6.2 Sol (over 272K) OpenAI $4.00 $15.00
GPT-6.2 Astra OpenAI $8.00 $40.00
GPT-6 Luna OpenAI $0.10 $0.50
Grok 4.7 (up to 200K) xAI $2.00 $6.00
Kimi K3.1 Pro Moonshot AI $2.00 $10.00
Qwen3.8-Max Alibaba (QwenCloud) $2.00 $6.00
GLM-5.3 Z.AI $1.40 $4.40
Muse Spark 1.3 Meta $1.25 $4.25
Muse Wonder 1.0 Meta $4.00 $20.00
DeepSeek-V4.1-Flash (peak hours) DeepSeek $0.30 $1.20
MiniMax-M3 MiniMax $0.30 $1.20

Prices compiled from VentureBeat's November 18, 2026 comparison and the vendors' pricing pages.

Capability Comparison

Model AA Intelligence Index Vals Index Strongest at
GPT-6.2 Astra 62 (per one source) - Top tier, very expensive
Gemini 4 Neon 61 70.9% Long context, automation, science, price/performance
Gemini 4 Carbon 60 70.5% Hardest long projects
Claude Opus 5.5 59 70.1% Terminal and agentic coding
GPT-6.2 Sol 57 66.1% Chart reading, cheap tokens
Claude Sonnet 5.5 55 70.85% Very balanced for its price
Gemini 4 Hydrogen 54 68.5% Speed and free tier

Brand By Brand

Anthropic Claude Opus 5.5. Neon's most direct rival. Neon is 25% cheaper per token and leads on most tests. Still, for terminal-based coding agents, sandbox security and long coding sessions, Opus 5.5 remains the pick of many developers. In one coding benchmark, the top spot goes to Anthropic's larger Claude Fable 5.2, with Neon second.

Anthropic Claude Sonnet 5.5. A third cheaper than Neon ($2 / $10) and almost identical on the Vals Index (70.85% vs 70.9%), but 6 points behind on the Intelligence Index. A very strong choice when budget comes first.

OpenAI GPT-6.2 Sol. Cheaper per token ($2 / $10), but rises to $4 / $15 once input passes 272K tokens, while Neon stays flat up to 1M. Neon leads on most tests; Sol is better at chart reading. VentureBeat estimates Neon needs about a third fewer tokens per task to offset its price premium.

OpenAI GPT-6.2 Astra and Claude Fable 5.2. Both are larger, slower and about 2.7x more expensive than Neon ($8 / $40). These are the "larger, slower models" ranked above Neon on the Artificial Analysis index. Valuable for the hardest research and coding work, often overkill for everyday business tasks.

xAI Grok 4.7. Very low output price ($2 / $6). It was not included in the same benchmark tables as Neon, so there is no reliable data on the capability gap. Worth testing for price-sensitive work.

DeepSeek, MiniMax, Xiaomi MiMo. These run at a tenth to a twentieth of Neon's price (for example DeepSeek-V4.1-Flash at $0.30 / $1.20, half that off-peak). For very high-volume, simple work they cut costs dramatically. Businesses should check where data is processed and whether that meets their privacy and compliance requirements.

Meta Muse, Moonshot Kimi K3.1, Qwen3.8-Max, GLM-5.3. Strong options in the mid and low price bands, and some can be self-hosted as open models. Because they are not in the same official benchmark tables as Neon, a direct capability comparison would not be fair.

Where to Use It

  • Gemini app: Google AI Pro ($19.99/month in the US) and AI Ultra (from $99.99/month) subscribers, listed as "Deeper analysis". Not on the free tier or AI Plus. Rollout is gradual.
  • Gemini API and Google AI Studio: Paid projects only. Model ID gemini-4-neon.
  • Google Cloud / Vertex AI: For enterprise customers in global, US and EU regions.
  • Google Antigravity: For AI Pro and Ultra subscribers. Also available in Android Studio.
  • OpenRouter: google/gemini-4-neon.
  • Not yet: As of November 18 it was not in Cursor or the Gemini CLI.

Safety

According to Google's own testing:

  • On Gray Swan's indirect prompt injection benchmark, no attack succeeded at the highest thinking level.
  • Its shortcut (cheating) rate on a coding eval is 2.45%, the lowest in the Gemini 4 family; 0% on Terminal-Bench 4.0.
  • 99.2% on an agent honesty eval. Compared with Carbon, under-service dropped 43-56% and overreach 15-35%.

All of this is Google's own measurement; independent verification is still limited.

Weaknesses and Criticism

  • Most results come from Google. Independent tests may differ, and early reactions depend heavily on which tasks people measure.
  • No official Neon vs Carbon comparison. The "Carbon quality" claim rests on index scores for now.
  • Pricier than some rivals. 50% more per token than Claude Sonnet 5.5 and GPT-6.2 Sol. It only pays off when it finishes the job with fewer tokens and fewer retries.
  • No free tier. Even trying it requires a paid API project or an AI Pro subscription.
  • No image or audio generation. You need other models for that.
  • Confusing names. A common complaint is that Hydrogen, Neon, Carbon and Argon say nothing about which model is bigger.

Which Model for Which Business Task?

Need Our pick
Analysing long contracts, reports, large document sets Gemini 4 Neon (1M context, flat price, best long-context results)
Spreadsheets, accounting, finance, workflow automation Gemini 4 Neon
Coding agent for a website or app Gemini 4 Neon or Claude Opus 5.5 (test both on your own work)
Customer service chat, fast replies, classification Gemini 4 Hydrogen or Claude Haiku 5.5
Budget-first, balanced general use Claude Sonnet 5.5 or GPT-6.2 Sol
Very high-volume simple text work DeepSeek, MiniMax, Gemini 3.5 Flash-Lite (check data location)
Hardest research and long projects GPT-6.2 Astra, Claude Fable 5.2, Gemini 4 Carbon

Our practical advice: don't lock into one model. Using Neon for planning and analysis and Hydrogen or a cheaper model for high-volume steps in the same workflow cuts costs significantly. Cache repeated instructions (95% off) and run non-urgent jobs through Batch at half price.

Conclusion

Gemini 4 Neon looks like Google's most balanced model so far: ahead of Claude Opus 5.5 and GPT-6.2 Sol on independent intelligence indexes, level with Carbon, about twice as fast, and 25% cheaper than Carbon from December 1. It leads clearly in long-document analysis, automation and science. Claude Opus 5.5 is still a strong rival for terminal-based coding, and on per-token price Sonnet 5.5 and GPT-6.2 Sol are cheaper. The right choice depends on the type and volume of your work; the best approach is a small trial with your own data.

If you want to add AI to your website, app or business workflow, Karakaş Digital can help you choose the right model, set it up and estimate the cost. Reach us at +90 850 241 92 50 or info@karakas.digital.

Sources