Skip to main content
株式会社オブライト
株式会社オブライト

Gemini 4 Neon Explained

$3/$15 Pricing, Benchmarks, API and Carbon vs Hydrogen

Published 2026/11/198 min read

Google Gemini 4 Neon (Nov 18, 2026) costs $3/$15 per 1M tokens, ID gemini-4-neon, 1M context. Covers Carbon vs Hydrogen, benchmarks, API code, cost math.

Gemini 4 Neon is a mid-tier Gemini 4 model Google announced on November 18, 2026, aimed at complex engineering and knowledge work where Hydrogen falls short but Carbon is overkill. The model ID is gemini-4-neon; pricing is $3 input / $15 output per 1M tokens, $0.15 cached input, with a 1M (1,048,576) token context at a flat rate. There is no free tier. It is available on the paid Gemini API, Google AI Studio, Google Antigravity and the Gemini app (Pro / Ultra).

- What it is: sits between Gemini 4 Hydrogen (fast, cheap) and Gemini 4 Carbon (top tier), for long, tool-heavy agent work and multi-file edits
- Cost: $3 / $15, twice Hydrogen and about 75% of Carbon's post-December 1 price ($4 / $20). Batch and Flex are half price; Priority is $5.40 / $27
- Reported results (Google): Artificial Analysis Intelligence Index 61 at xhigh (#3), Terminal-Bench 4.1 65.2%, HLE-Diamond 71.3%; Google says it costs 32% less per task than Carbon at high

This article relies on Google's blog post and the Gemini API pricing, model, "What's new" and "Choosing a model" pages. All benchmark numbers are Google's published figures and were not reproduced by us.

Gemini 4 Neon specs at a glance

ItemDetails
AnnouncedNovember 18, 2026 (Google blog "Introducing Gemini 4 Neon")
PositionBetween Carbon (hardest long-horizon work) and Hydrogen (fast everyday execution)
Model IDgemini-4-neon
Context1,048,576 tokens, shared by input, thinking and output
Max output65,536 per synchronous call; longer with background: true and max_output_tokens
Input / outputText, image, video, audio, PDF in; text only out (no image/audio generation, no Live API)
Thinking levelslow / medium / high (default) / xhigh; minimal returns an error
CachingImplicit only; explicit caching is not supported
Knowledge cutoffNot stated on the official pages

Positioning: Carbon, Neon and Hydrogen

Google's "Choosing a model" guide calls Hydrogen the starting point for most projects, Neon the step up when harder reasoning is needed without Carbon's extra verification, and Carbon the choice for architecture, difficult debugging, proofs and security audits. Neon defaults to high and supports xhigh like Hydrogen, whereas Carbon and Argon default to medium and top out at high, so swapping Carbon for Neon changes the default reasoning depth.

ModelBest for (per Google)SpeedInput / output per 1MFree tier
Gemini 4 HydrogenEveryday coding, interactive apps, worker agentsFast$1.50 / $7.50Yes
Gemini 4 NeonComplex engineering, dense knowledge work, scienceModerate$3 / $15No
Gemini 4 CarbonArchitecture, hard debugging, proofs, security auditsDeliberate$2 / $10 through Nov 30; $4 / $20 from Dec 1No

Pricing and monthly estimates

TierInputOutput (incl. thinking)Cached input
Standard$3.00$15.00$0.15
Batch / Flex$1.50$7.50$0.075
Priority$5.40$27.00$0.27

The pricing page lists no over-200K surcharge and no scheduled price change for Neon. Thinking tokens, and internal system instructions and function declarations used with grounding, are billed. Google Search grounding (Standard and Priority only) includes 5,000 free requests a month, then $14 per 1,000.

ModelMonthly cost (10M input, 2M output)Note
Gemini 4 Hydrogen$30
Gemini 4 Neon (Standard)$60about 9,000 JPY at 150 JPY/USD
Gemini 4 Neon (Batch / Flex)$30for non-urgent work
Gemini 4 Neon (70% of input cached)about $40
Gemini 4 Carbon (from Dec 1)$80$40 until Nov 30

These are unit prices multiplied out; real bills depend on thinking tokens, and heavy xhigh use raises output volume.

Benchmarks reported by Google

AreaBenchmarkNeonGoogle's note
GeneralArtificial Analysis Intelligence Index61 (xhigh)#3 overall, behind larger, slower models
GeneralVals Index70.9%#2
AgentsAgent Arena Index1336#2
AgentsAgents' Last Exam v1.175.8%on par with leading models
AgentsZapier AutomationBench56.3%#1
ScienceTerminal-Bench Science / RiemannBench / HLE-Diamond70.5% / 84.0% / 71.3%best in class
CodingAA Coding Agent Index58.6#1
CodingTerminal-Bench 4.1 / FrontierCode 2.0 (high)65.2% / 46.2%
CyberSEC-bench Pro / WebExploitBench v1.2 / CWE-bench v290.5% / 51.26% / 70.7%state of the art; CWE-bench +3.4% over Carbon

Compared with the figures in our Hydrogen article, Neon moves the Intelligence Index from 51 to 61, Terminal-Bench 4.1 from 57% to 65.2%, HLE-Diamond from 58.8% to 71.3% and Agents' Last Exam from 71.6% to 75.8%. Google also claims 32% lower cost per task than Carbon at high, and 50-80% lower than top competing models on Vals Index and Agent Arena. These are task-level figures, not unit prices, and mostly use unnamed competitors.

Availability

WhereStatus (as of Nov 18, 2026)
Gemini APIPaid tier only, gemini-4-neon
Google AI StudioAvailable; Pro / Ultra subscribers can try it
Google AntigravityPaid Pro / Ultra; standalone or orchestrating Hydrogen subagents
Gemini appPro / Ultra via the model dropdown
Gemini EnterpriseVia API and model dropdown
Neon for cyberRolling out to trusted defenders in the Fairwind Program
Vertex AINot mentioned on the pages we checked

Using the Gemini API

from google import genai

client = genai.Client()  # uses GEMINI_API_KEY
response = client.models.generate_content(
    model="gemini-4-neon",
    contents="Find the cause of the auth bug in this repo and propose a fix",
)
print(response.text)

Google's official quickstart uses the Interactions API instead, with the thinking level set in generation_config.

from google import genai

client = genai.Client()
interaction = client.interactions.create(
    model="gemini-4-neon",
    input="List the differences between this spec and the spreadsheet",
    generation_config={"thinking_level": "high"},  # low / medium / high (default) / xhigh
)
print(interaction.output_text)

- Drop temperature, top_p and top_k: in-range values are ignored, out-of-range values return 400
- minimal thinking is rejected
- max_output_tokens caps thinking plus output; remove low caps such as 8,192
- No explicit caching; implicit caching applies from 4,096 tokens
- Raise client timeouts (15 minutes or background execution) for hard prompts
- Migrating from Carbon, Argon or Hydrogen only requires changing the model ID

Comparison with other models

Other vendors' figures come from our existing articles on Claude Fable 5.2 and GPT-6.2 Astra / Sol.

ModelInput / output per 1MMonthly (10M in, 2M out)
Gemini 4 Hydrogen$1.50 / $7.50$30
GPT-6.2 Sol$2 / $10$40
Gemini 4 Neon$3 / $15$60
Gemini 4 Carbon (from Dec 1)$4 / $20$80
Claude Opus 5.5$4 / $20$80
GPT-6.2 Astra$8 / $40$160
Claude Fable 5.2$8 / $40$160

Published scores are close in places (for example FrontierCode 2.0: Fable 5.2 at 49.1% vs Neon at 46.2%), but versions, harnesses and thinking settings may differ, so treat them as vendor-reported figures rather than a head-to-head result.

How to choose

SituationModelReason
Everyday coding, worker agentsHydrogenOfficial starting point, free tier, half the price
Complex work where Hydrogen falls shortNeonClearly higher reported scores; cheaper per task than Carbon per Google
Architecture, hard debugging, security auditsCarbonPositioned for the hardest problems
Bulk non-urgent processingNeon Batch / FlexHalf price

A practical path: run your own tasks on Hydrogen at high, rerun only the failures on Neon, then try xhigh or Carbon for what remains. Not published: knowledge cutoff, Vertex AI availability, exact start of the cyber variant, and the names of the anonymous competitor models.

FAQ

How much does Gemini 4 Neon cost?

$3 input / $15 output per 1M tokens (thinking included), $0.15 cached input. Batch and Flex are $1.50 / $7.50 and Priority is $5.40 / $27. The price is flat across the 1M context and there is no free tier.

How does Neon differ from Carbon and Hydrogen?

Google places Neon between Hydrogen (fast everyday execution) and Carbon (hardest long-horizon tasks). It costs twice Hydrogen and about 75% of Carbon's post-December 1 price. Neon defaults to high and supports xhigh; Carbon defaults to medium and tops out at high.

What is the model ID and how do I call it?

The ID is gemini-4-neon. With the google-genai Python SDK you can call client.models.generate_content(model="gemini-4-neon", contents=...), or follow Google's quickstart with client.interactions.create and thinking_level in generation_config. Remove temperature, top_p and top_k.

Can I try it for free?

Not through the free API tier. Paid API projects, and Google AI Pro / Ultra subscribers in AI Studio, Antigravity and the Gemini app, can use it. Hydrogen has a free tier.

Are the benchmark numbers reliable?

They are Google's published figures plus third-party results Google cites; we did not reproduce them, and most competitors are unnamed. Test 10 to 30 of your own tasks for quality and cost before adopting it.

株式会社オブライト

Oflight Inc. is a Tokyo-based IT company providing software development, digital signage, web development, network systems, and IoT/AR/VR technology services.

Related free tools (no sign-up, instant results)

Feel free to contact us

Contact Us