Google Launches Gemini 4 Neon for Coding and Analysis | Zeniteq
Loading
Zeniteq is now available on mobile for iPhone, iPad and Apple Watch.Download nowLearn more
Google Launches Gemini 4 Neon for Coding and Analysis
Google’s new model sits between Carbon and Hydrogen, with paid access, four thinking levels, and a shared token budget that changes how developers should plan requests.
Google launched Gemini 4 Neon on November 18, 2026, positioning it between Carbon and Hydrogen for complex, high-frequency coding and enterprise analysis workloads, with access through the paid Gemini API, Gemini for business, Google Antigravity, and the Gemini app.
Priced at $3 per million input tokens and $15 per million output tokens, Neon offers four thinking levels defaulting to HIGH, and Google claims 32% lower cost per task than Carbon at that level. Its 1,048,576-token budget is shared across input, reasoning, tool calls, and output combined — a constraint developers must account for before migrating existing workflows.
Google introduced Gemini 4 Neon on November 18, 2026, positioning it between Carbon’s long-horizon reasoning and Hydrogen’s faster everyday execution. The model is rolling out through the paid-tier Gemini API, Gemini for business, Google Antigravity, and the Gemini app.
Google describes Neon as delivering frontier-level coding and analytical performance with faster task completion. Its practical appeal rests on more specific details: $3 per million input tokens, $15 per million output tokens, four thinking levels, and a 1,048,576-token budget shared by input and output. Those constraints matter when moving an existing coding agent or document-analysis workflow onto the model. (Google announcement)
The shared budget deserves particular attention. A million-token context window is not a promise that developers can submit a million tokens and then generate another million. Neon’s input, reasoning, function calls, and final answer must fit within the same allowance.
Where Neon Fits Between Carbon and Hydrogen
Google’s November 18 model card says Neon is adapted from Gemini 4 Carbon. The company describes the family’s division of labor this way:
Model
Google’s intended role
Gemini 4 Carbon
The hardest, long-horizon projects
Gemini 4 Neon
Complex, high-frequency coding, analysis, and enterprise workflows
Gemini 4 Hydrogen
Fast, everyday execution
Neon can also act as an orchestrator. Google’s developer guide describes using gemini-4-neon to plan and verify work, then passing previous_interaction_id to gemini-4-hydrogen subagents through the Interactions API for lower-latency execution. (Developer guide)
That makes Neon a candidate for the middle of an agent stack: repository-level planning and difficult decisions on Neon, with narrower execution tasks delegated to Hydrogen. That is an architectural implication of Google’s documentation, not a configuration tested for this article.
I’d evaluate Neon against Carbon using completed-task cost, including retries and tool calls. Google reports 32% lower cost per task than Carbon at Neon’s default HIGH thinking level in its benchmark comparison. That is a vendor-reported result under evaluation conditions, not a guaranteed saving for every codebase. (Google announcement)
Where Developers Can Access Neon
Google’s launch lists these access routes:
Gemini API through Google AI Studio: paid-tier API access.
Gemini for business: API access and the model dropdown.
Google Antigravity: paid Google AI Pro and Ultra subscribers.
Gemini app: Pro and Ultra subscribers.
These are separate product surfaces. The published per-token API rates should not be read as an explanation of app subscription quotas or Antigravity allowances. (Google announcement)
The Enterprise Agent Platform route has an additional restriction: its guide says Neon is available to verified enterprise customers in select regions. Developers should check that route’s eligibility before planning a deployment around it. (Developer guide)
For cyber work, treat Fairwind access as a separate rollout to verify, not as something automatically unlocked by general Neon API or app access. The general-access details cited here do not specify Fairwind eligibility or timing.
Neon accepts text, images, audio, and video. The enterprise guide also documents direct Microsoft Office and OpenDocument inputs, including Word documents, presentations, and spreadsheets. That creates a useful path for analyzing mixed business material without first converting every file to plain text. Its model card lists a March 2026 knowledge cutoff and warns that hallucinations remain possible, so current factual work still needs supplied evidence or retrieval. (Developer guide, model card)
The Million-Token Budget Is Shared
The developer guide defines Neon’s limit as 1,048,576 tokens across input and output combined. Output includes thinking, function calls, and the final response.
A request containing 900,000 input tokens therefore leaves at most 148,576 tokens for everything the model generates. Filling the entire budget with input would leave no generation space. A long reasoning process also reduces the room available for the answer the user actually sees. (Developer guide)
My practical recommendation is to reserve output space before assembling a large prompt. A repository that technically fits inside the context window may still leave too little room for analysis, tool use, and a substantial patch.
Long Output Requires Continuation
Google’s model card describes a default synchronous output allowance of 64K tokens. The developer guide supplies the more precise operational detail: a foreground segment is limited to 65,536 output tokens, with longer generation continuing across segments.
The documented routes are:
Automatic continuation handled by the Google Gen AI SDK.
Server-managed asynchronous generation using background=true.
A custom loop that handles continuation_token.
The guide documents long generation through both the Interactions API and GenerateContent v1beta1, up to the unified limit minus input tokens. The distinction is between a single foreground segment and the total generation budget. An application consuming very long results needs continuation-aware request handling, timeouts, and output storage. (Developer guide)
The following are calculations from Google’s published rates, not measured bills. They assume uncached input and output charged at the listed output rate, including the thinking and function-call tokens counted as output in the guide. (Pricing announcement, token definitions)
Input tokens
Output tokens
Combined tokens
Calculated token cost
100,000
20,000
120,000
$0.60
500,000
500,000
1,000,000
$9.00
500,000
548,576
1,048,576
$9.73
0, theoretical boundary
1,048,576
1,048,576
$15.73
The output-only row is a mathematical boundary, not a realistic prompt. Its exact calculation is 1,048,576 × $15 / 1,000,000 = $15.72864.
Because output costs five times as much as standard input, shifting a fixed budget toward generation raises the bill. At these rates, about $15.73 is the theoretical token-charge maximum for one fully occupied unified budget, excluding separate service charges. It is not a spending cap for an agent session, which can make multiple billable calls.
Caching Has a Specific Constraint
Neon’s cached-input rate is 95% below its standard input rate, but developers cannot assume an existing explicit-cache integration will transfer unchanged.
The guide says explicit context caching is unsupported. Implicit caching applies automatically to shared prompt prefixes of at least 4,096 tokens, with durable caching enabled by default. I’d keep stable instructions and reusable context consistently ordered at the front of prompts, while checking actual cache usage before budgeting around the discount. (Developer guide)
Four Thinking Levels Control Reasoning Depth
Neon accepts LOW, MEDIUM, HIGH, and XHIGH, with HIGH as the default. Google describes XHIGH as adding deeper, multistep reasoning for the hardest engineering, analytical, and scientific workloads. (Google announcement)
I’d begin a migration at HIGH, establish a task-success baseline, then try lower settings on bounded work. Escalating difficult cases to XHIGH makes more sense than enabling it universally, particularly because thinking consumes the same output budget as the final answer.
Google’s documented Python request pattern sets the level inside ThinkingConfig:
from google import genai
from google.genai import types
client = genai.Client(
enterprise=True,
project="PROJECT_ID",
location="global",
)
response = client.models.generate_content(
model="gemini-4-neon",
contents="Review this design for concurrency risks.",
config=types.GenerateContentConfig(
thinking_config=types.ThinkingConfig(
thinking_level="HIGH",
),
),
)
print(response.text)
This example shows request structure; it has not been executed here. Replace PROJECT_ID, configure Application Default Credentials or an API key as documented, and use an eligible project. (Developer guide)
Migration Requires More Than a Model-ID Swap
Google’s guide covers migration from Gemini 3.1 Pro, Gemini 4 Argon, and Gemini 3.8 Flash. That documents a migration path, not a retirement announcement for those models.
SDK versions matter if the application relies on automatic continuation:
API
Python minimum
JavaScript/TypeScript minimum
Interactions API
google-genai >= 2.34.0
@google/genai >= 2.32.0
GenerateContent
google-genai >= 2.30.1
@google/genai >= 2.29.1
The guide also lists minimum versions for other supported languages. (Developer guide)
Before switching production traffic, check:
Budget handling: reserve space for thinking, tool calls, and answers.
Long-response handling: support continuation and interrupted or partial results.
Caching: remove assumptions that explicit context caching is available.
Thinking settings: set the intended level explicitly.
Access and terms: verify regional eligibility and the Google Frontier Model conditions referenced under Section 34.
These checks are more consequential than changing a model string. An integration can send valid requests while still mishandling long responses or underestimating output spend.
Benchmarks Show Why the Agent Harness Matters
Google’s announcement reports an Artificial Analysis Intelligence Index score of 61 at XHIGH, placing Neon third overall, and a Vals Index score of 70.9%, placing it second. It also reports first place on the Artificial Analysis Coding Agent Index with 58.6, alongside 65.2% on Terminal-Bench 4.1 and 46.2% on FrontierCode 2.0 at HIGH.
Some are third-party evaluations, but the evidence here is Google’s presentation of their results, not independent testing conducted for this article. (Google announcement)
The most useful comparison may be within Neon itself. Google reports 12.8% on SWE-sweep using the mini-SWE harness, rising to 39.0% in Google Antigravity. The published score is roughly three times higher when the surrounding agent environment changes. (Model card)
That gap is a reason to evaluate Neon inside the workflow you intend to ship. Tool access, orchestration, context management, and retry behavior belong in the comparison alongside the model. The same distinction matters when choosing among agentic AI coding tools: a strong model score cannot describe the whole development experience.
Frequently Asked Questions
4 questions
1
Is Gemini 4 Neon Available for Free?
The launch specifies paid-tier API access and Pro or Ultra access in the Gemini app and Antigravity. It does not announce free-tier availability.
2
Can Neon Generate a Million Tokens in One Response?
It supports long generation up to the remaining unified budget. Foreground segments are limited to 65,536 output tokens, so larger results require continuation or background generation.
3
How Much Does a Million-Token Workload Cost?
At standard rates, 500,000 input tokens plus 500,000 output tokens cost $9. The total depends on the input/output split and whether input receives cached pricing.
4
Which Thinking Level Is the Default?
HIGH is the default. The other supported levels are LOW, MEDIUM, and XHIGH.