What Is Gemini 4 Neon? Google’s Frontier-Efficiency Model Explained (2026)

Ethan Coleon 2 hours ago

Short answer: Gemini 4 Neon is Google’s new Gemini 4 model for complex coding, knowledge work, and science workflows, announced on 18 November 2026. Google positions it between Carbon, for the hardest long-horizon projects, and Hydrogen, for fast everyday execution. The official API model code is gemini-4-neon. Google lists a 1,048,576-token unified context window, multimodal input, and text output. Standard API pricing is $3 per million input tokens, $0.15 cached input, and $15 output, including thinking tokens in the output price.

The important qualifier is the product surface: the API’s 1M context is not a promise that every app exposes the same limit. Google’s Antigravity announcement separately lists up to 512K there. This explainer separates those figures, rollout details, and vendor benchmark claims rather than treating them as one universal spec.

Gemini 4 Neon at a glance

QuestionConfirmed answer
What is it?A Gemini 4 model balancing reasoning depth and efficiency for frequent, complex work
Announced18 November 2026, in Google’s launch post
API model IDgemini-4-neon, listed as stable
API context1,048,576 tokens, unified across input and output
API output65,536 tokens for default synchronous use; up to 1,048,576 with long decode, subject to the unified context budget
Input / outputText, images, video, audio, and PDF in; text out
Standard token prices$3 input / $0.15 cached input / $15 output per million tokens
API free tierNot available for Neon, according to the pricing page
Other rollout surfacesGemini app, Google Antigravity, Gemini for business, and paid-tier Gemini API in AI Studio
Our testingSource-based introduction, not an independent benchmark or hands-on performance test

Where Neon fits: Carbon, Neon, and Hydrogen

Google’s family description is useful because “newer” does not automatically mean “replace every other model.” In the launch post, Carbon remains the choice for the hardest, long-running projects. Hydrogen handles fast everyday execution. Neon is aimed at complex, high-frequency workflows that need deeper reasoning without always paying the time and task cost of a larger model.

  • Carbon
    • Google’s positioning: Hardest, long-horizon projects
    • A sensible evaluation starting point: A difficult investigation or architecture problem where depth matters most
  • Neon
    • Google’s positioning: Complex coding, analytical reasoning, production enterprise workflows
    • A sensible evaluation starting point: Repeated repository edits, document analysis, and tool-heavy work that needs both judgment and throughput
  • Hydrogen
    • Google’s positioning: Fast everyday execution
    • A sensible evaluation starting point: Routine implementation and rapid iteration

These are positioning statements, not guarantees that Neon will win your workload. Compare them on the same task, with the same tools, success criteria, and reasoning effort. For the earlier flagship, see What is Gemini 4 Carbon?.

Neon as an orchestrator

Google also describes Orchestrator (Neon) in Antigravity. Here, Neon plans and reviews while Hydrogen performs implementation work such as edits and test runs. If Hydrogen gets stuck, control returns to Neon. That is a product workflow using two models, not proof that a single plain API call automatically starts a swarm.

The Google staff announcement lists separate access rules for Neon and Orchestrator. In particular, enterprise Neon access through Gemini for business does not imply Orchestrator access.

Official capabilities and limits

The Gemini API model page confirms a stable model ID and a broad multimodal input stack. It accepts text, images, video, audio, and PDFs, and returns text. Image generation, audio generation, and the Live API are listed as unsupported. Understanding a video or an image is therefore not the same thing as generating one.

The API page lists support for code execution, computer use, file search, function calling, structured outputs, URL context, and grounding with Google Search and Maps. It also lists Batch, Flex, and Priority consumption options. Those capabilities still depend on how the integration exposes tools and permissions; a model name alone does not authorize actions.

Context and output are different budgets

The official model page calls the 1,048,576-token context window unified: input and output share that budget. It gives 65,536 as the default synchronous maximum output and says long decode can reach 1,048,576. Do not add a full million-token input to a full million-token output and assume both fit together.

Antigravity is a separate surface. Google’s staff announcement says up to 512K tokens in Antigravity. The launch’s headline “1 million-token context” and that app limit are not contradictory when kept in their proper scopes.

Thinking levels

The launch post describes four thinking levels, with high as the default and xhigh for deeper multi-step work. The Antigravity announcement names Low, Medium, High, and Extra, and shows xhigh as the CLI effort value. More effort can change latency, token use, and task outcomes. A score measured at xhigh is not automatically the score of a default-high session.

Pricing: what the $3 / $15 headline leaves out

The official pricing page lists Neon as a paid-tier model:

Standard pricing itemUSD per 1M tokens
Input$3.00
Cached input$0.15
Output, including thinking tokens$15.00

For a simple illustration, 100,000 uncached input tokens plus 10,000 billed output tokens would cost $0.45 at these token rates: $0.30 input and $0.15 output. That illustration excludes grounding charges and other tools, and the output budget must account for thinking tokens, not just the text you see.

Caching is implicit only; explicit caching is not supported. The $0.15 rate is not a promise that every repeated prompt receives a cache hit. Google’s pricing page also warns that model-processed tokens include internal tool instructions and function declarations when using grounding.

Search grounding has a separate allowance and billing line: the page lists 5,000 free monthly search requests shared across Gemini 3 and newer models, then $14 per 1,000 requests, with additional query-count and daily-cap notes. Check that page for your exact usage rather than budgeting from token price alone.

Finally, cost per task is not price per token. A model can finish with fewer tokens or fewer turns and lower the final bill even if its token rate is higher than a temporary promotional rate on another model. Google’s efficiency claims should not be read as a universal discount on every request.

Where can you use Gemini 4 Neon?

Google describes a rollout beginning on 18 November, not a guarantee that every account sees every option immediately.

  • Developers: the Gemini API paid tier in Google AI Studio and Android Studio. The launch post also says Pro and Ultra subscribers can try Neon in AI Studio.
  • Consumers: Google AI Pro and Ultra subscribers can select Neon in the Gemini app’s model dropdown.
  • Antigravity: the launch targets paid Google AI Pro and Ultra subscribers. The staff post covers the app, CLI, and IDE extensions; it excludes Free and Google AI Plus and limits Pro-trial eligibility to trials with eligible premium devices.
  • Enterprises: Gemini for business via API or model dropdown, plus Antigravity access as described by Google.
  • Cybersecurity defenders: the separate Neon for cyber is rolling out to eligible Fairwind Program members in the coming days. That is not the same access promise as the general model.

If Neon is absent, first check the product, plan, and rollout status. Do not assume that a paid API project, a consumer subscription, and an enterprise workspace are interchangeable entitlements.

What the launch benchmarks actually say

Google reports strong results across coding, knowledge work, and science. Selected numbers in its launch post include:

EvaluationGoogle’s reported resultCaveat
Artificial Analysis Intelligence Index61, #3 overallAt xhigh thinking; a launch-time ranking, not a permanent position
Artificial Analysis Coding Agent Index58.6, #1 overallReported by Google in the announcement
Agents’ Last Exam v1.175.8%A specific benchmark setup, not a pass rate for all office work
Terminal-Bench 4.165.2%Tooling and harness matter
FrontierCode 2.046.2%At the default high setting

Google also reports about 32% lower cost per task than Carbon at Neon’s default high thinking level. We have not reproduced these measurements. The post itself demonstrates why harness context matters: it reports 12.8% on SWE-sweep with the mini-SWE harness versus 39.0% in Antigravity. Neither number should be detached from that environment.

For a buying or deployment decision, use a small representative task set: one familiar code fix, one document with a known answer, and one multi-step workflow. Check correctness, review time, tool permissions, elapsed time, and the complete bill. Those are suggested evaluation criteria, not tests we claim to have run.

Safety and practical limitations

Google says Neon builds on the Gemini 4 safety stack and maintains Carbon-level protections against severe misuse and prompt injection. Its reported honesty and injection-resistance results are bounded evaluations, not claims that it never fabricates or that every external document is safe to follow.

Treat generated work as reviewable output. Verify calculations and citations, inspect code changes, and require appropriate approval before an agent changes external systems. A low internal fabrication rate does not remove the need for those checks.

This article is a model introduction, not an adult-content or jailbreak guide. For a separate creative-chat product, HackAIGC Chat is another destination to explore; that link is not a claim that HackAIGC runs Gemini 4 Neon or matches Google’s specifications.

FAQ

What is Gemini 4 Neon?

Google’s November 2026 Gemini 4 model balancing reasoning depth and task efficiency for complex coding, knowledge work, and science. The official stable API ID is gemini-4-neon.

Is Neon better than Carbon?

Not universally. Google positions Carbon for the hardest long-horizon projects and Neon for frequent complex workflows with better efficiency. Match the model and thinking effort to your own tasks.

Does Neon have a 1M-token context?

The API lists a unified 1,048,576-token input/output budget. Antigravity separately documents up to 512K. Default synchronous output is 65,536; larger long-decode output is not an extra budget beyond the unified context.

Is Gemini 4 Neon free?

The API pricing page says Neon is not available on the Free tier. Consumer and Antigravity access depend on the supported plan and rollout.

Can Neon generate images or audio?

The API model page lists text output, with image and audio generation unsupported. Multimodal input is not multimodal output.

Did you independently test the model?

No. This is a source-verified introduction. Performance figures are attributed to Google’s launch post; no hands-on scores or availability tests are invented.