
Automatically append previous messages to maintain multi-turn context. May increase token usage.
How can I help you today?
Automatically append previous messages to maintain multi-turn context. May increase token usage.
How can I help you today?
README
Complete guide to using Gemini 4 Neon
Gemini 4 Neon API for Complex Coding, Analysis and Science

curl https://api.kie.ai/gemini-4-neon-openai/v1/chat/completions \-H "Authorization: Bearer $KIE_API_KEY" \-H "Content-Type: application/json" \-d '{"messages": [{ "role": "user", "content": "Analyze the latency vs consistency trade-offs of multi-region consensus." }],"reasoning_effort": "high","stream": false}'
Key Capabilities of Gemini 4 Neon


Repository-Level Coding and Multi-File Refactoring

Four Thinking Levels From Low to Xhigh

Unified 1M-Token Multimodal Context Window

Analytical, Legal and Scientific Workflows

Secure Code Generation, Grounding and Tools
Gemini 4 Neon Specifications
Model specifications and Kie.ai API parameters for Gemini 4 Neon from Google's model card and the Kie.ai API documentation.
| Specification | Gemini 4 Neon |
|---|---|
| Developer | Google (Google DeepMind) |
| Model IDs on Kie.ai | gemini-4-neon-openai (OpenAI-compatible); gemini-4-neon (native Gemini) |
| Release date | November 18, 2026 |
| Adapted from | Gemini 4 Carbon |
| Modalities | Text, image, video, audio and PDF input (up to 10 files per request); text output |
| Context window | 1,048,576 tokens unified across input, thinking and output |
| Max output tokens | Up to 65,536 tokens per synchronous response by default (thinking tokens count toward output) |
| Thinking levels | low, medium, high (default), xhigh via reasoning_effort or thinkingLevel (minimal ignored; include_thoughts / includeThoughts optional) |
| API formats on Kie.ai | POST /gemini-4-neon-openai/v1/chat/completions (stream default false); POST /gemini/v1/models/gemini-4-neon:streamGenerateContent (stream default true) |
| Tools & parameters on Kie.ai | Google Search grounding (googleSearch), function tools, response_format JSON schema (OpenAI-compatible), implicit caching (temperature, topP and topK have no effect; chat requests only) |
| Knowledge cutoff | March 2026 |
Gemini 4 Neon Benchmark Results
Evaluation results reported by Google in its November 18, 2026 release announcement and DeepMind model card.
| Benchmark | Gemini 4 Neon | What It Measures |
|---|---|---|
| AA Intelligence Index v5.0.1 | 61 (#3 overall, at xhigh) | General capability index |
| Vals Index v2.2 | 70.9% (#2 overall; #1 Legal Research & Tax Agent) | Enterprise, legal and tax evaluations |
| AA Coding Agent Index | 58.6 (#1 overall) | Software engineering capability evaluation |
| CursorBench 4.1 / Terminal-Bench 4.1 | 65.4% / 65.2% (cheating-adjusted) | Repository coding and terminal execution |
| FrontierCode 2.0 / SWE-sweep | 46.2% (at high) / 39.0% (Antigravity harness), 12.8% (mini-SWE) | Software engineering benchmarks |
| Agents' Last Exam v1.1 / AutomationBench | 75.8% / 56.3% (#1 overall) | Multi-step knowledge work and workflow automation |
| RiemannBench / HLE-Diamond / Terminal-Bench-Science 0.1 | 84.0% / 71.3% / 70.5% | Mathematics and scientific reasoning |
| GraphWalks (256K to 1M, BFS F1) | 95.2% | Long-context retrieval and reasoning |
| LVBench / Chartography / MKA | 89.0% / 73.5% / 56.5% | Long video, dense charts and multimodal reasoning |
| SEC-bench Pro / CWE-bench v2 / WebExploitBench v1.2 | 90.5% / 70.7% / 51.3% | Vulnerability detection, patching and web security |
How to Work With Gemini 4 Neon
- 1
Step 1: Choose Between Hydrogen, Neon and Carbon
Use Gemini 4 Neon for complex software engineering, analytical reasoning and science workflows where you want frontier-level reasoning depth with efficient end-to-end task completion. Keep Gemini 4 Hydrogen for fast everyday coding and execution, and reserve Gemini 4 Carbon for the hardest long-horizon projects. - 2
Step 2: Start at High and Reserve Xhigh for Hard Problems
High is the default thinking level on Kie.ai. Set reasoning_effort (OpenAI-compatible) or generationConfig.thinkingConfig.thinkingLevel (native) to medium or low for leaner token use on simpler prompts, or xhigh for hard multi-step engineering, analytical and scientific problems. Leave temperature, topP and topK unset, as they have no effect on Gemini 4. - 3
Step 3: Leave Room for Thinking and Output Tokens
Input, thinking and output tokens share the 1,048,576-token window, and the default 65,536-token output budget covers both thinking and visible answer text. For hard xhigh tasks, remember that deeper thinking consumes more of that shared output budget, and pass up to 10 media files per request. - 4
Step 4: Use Grounding, Function Tools and Prefix Caching
Enable googleSearch when a question needs post-March 2026 web facts, declare custom function tools on either endpoint, or pass a response_format JSON schema in the OpenAI-compatible endpoint. Keep shared prompt prefixes unchanged across turns so repeated context can qualify for cached-input pricing.