Model Selection Guide

This document will help you understand the most suitable models to use in different scenarios in NekroAgent, and provide detailed performance, price, and applicability analysis. Currently, it mainly provides model selection information supplied by NekroAgent Official Relay, and will gradually add models from other sources.

Rating Description

In the recommended models, we use the following rating standards:

Rating	Corresponding Level	Description
👑	⭐⭐⭐⭐⭐	Excellent
🥇	⭐⭐⭐⭐	Outstanding
🥈	⭐⭐⭐	Good
🥉	⭐⭐	Average
⚪	⭐	Poor

Note

The following recommendations are for reference only. The same model from different sources may have differences in final performance due to channel conversion strategies, different configuration settings, concurrency situations, current status, etc. We encourage you to try multiple models based on actual usage, including those not in the following form, to choose the model that best suits you!

The models in the following tables are from NekroAgent Official Relay - Available Model List. If you think there is a significant difference between the following tables and actual experience, you are welcome to contact us for feedback. We will continuously maintain and update the tables to better match actual experience

For information on deprecated & discontinued models, please see Model Deprecations

NekroAgent Main Application

Chat Conversation Process

The chat session process of NekroAgent (excluding plugin functions) is mainly affected by three configuration items: Main Model Group (USE_MODEL_GROUP), Debug/Agent Migration Model Group (DEBUG_MIGRATION_MODEL_GROUP), and Fallback Model Group (FALLBACK_MODEL_GROUP). The specific scheduling strategy is as follows:

1. When a conversation process starts, the model in the Main Model Group is first used for generation
2. When the code generated by the Main Model Group triggers Agent type methods or produces program errors, subsequent calls in this process all use the model in the Debug/Agent Migration Model Group for iteration
3. If either the Main Model Group or Debug/Agent Migration Model Group model call fails, the model in the Fallback Model Group is used for generation
4. If the Fallback Model Group also fails to call, the response process ends in failure

Below is the list of recommended models for Chat Conversation Process:

This list updated on April 18, 2026

Model Name	Quality	Speed	Stability	Cost-Effectiveness	Vision	Built-in Thinking	Notes
claude-4-5-sonnet-latest	👑	🥈	🥈	🥈	👁️	❌	Anthropic's latest flagship model, with the strongest comprehensive capabilities but limited supply, suitable as the main model
gemini-3.1-pro-preview	👑	🥈	🥇	🥉	👁️	🧠	Google's 3.1 generation high-quality flagship model, currently top performance, supports thinking signature and thinking levels ⚠️ Preview model
gemini-3-flash-preview	🥇	🥇	🥇	👑	👁️	❌	Balanced model with excellent overall experience, fast speed and strong logic, recommended as the main model
gemini-2.5-pro	🥇	🥇	🥇	🥈	👁️	🧠	Stable logical ability, has adaptive thinking ability ⚠️ Expected to be discontinued on June 17, 2026
gpt-4.1	🥇	🥈	🥇	🥈	👁️	❌	Newer flagship GPT model, with obvious AI characteristics but decent logical ability
gemini-3.1-flash-lite-preview	🥈	👑	🥇	👑	👁️	❌	Ultra-fast small model, extremely low inference cost, suitable for simple tasks or fast iteration ⚠️ Preview model
claude-4-5-haiku	🥈	🥇	🥇	🥉	👁️	❌	Anthropic's fast model, suitable for scenarios with specific requirements for generation style
gemini-2.5-flash	🥇	🥇	🥇	👑	👁️	❌	High cost-effectiveness, will soon be replaced by gemini-3-flash ⚠️ Expected to be discontinued on June 17, 2026
deepseek-chat (v3)	🥇	🥉	🥇	🥈	❌	❌	Classic domestic model, excellent Chinese ability, distinctive language style
doubao-1.5-vision-pro-32k-250115	🥈	🥈	👑	🥈	👁️	❌	Domestic model provided by ByteDance, excellent stability, strong multimodal ability, suitable as a backup model
gemini-2.0-flash	🥈	👑	🥇	🥇	👁️	❌	Small model with extremely low cost ⚠️ Expected to be discontinued on June 1, 2026
gpt-4o	🥇	🥈	🥇	🥈	👁️	❌	Commonly used model for productivity scenarios, high API stability
gpt-4o-mini	🥈	🥈	🥇	🥇	👁️	❌	Classic GPT series small model
grok-3	🥈	🥈	🥇	🥉	👁️	❌	Language model launched by xAI, distinct personality, lower AI flavor

Note:

• In NekroAgent, the External Chain of Thought switch of the model first used in the conversation process (usually the main model) will affect the use of chain of thought in subsequent calls of this conversation process. For example, if the main model enables External Chain of Thought, the iteration/debug model will also have the effect of enabling external chain of thought
• Generally, models that support Built-in Thinking are not recommended to enable External Chain of Thought, otherwise it may reduce model generation speed
• Due to the implementation of the prompt iteration mechanism, it is not recommended to mix models that support vision and do not support vision, otherwise it may lead to request format errors

Plugin Development

The generation modification suggestion model in NekroAgent's Plugin Editor uses the Plugin Code Generation Model Group (PLUGIN_GENERATE_MODEL_GROUP) to generate code solutions for user needs. It is recommended to use models with strong coding capabilities and high quality. Below is the list of recommended models:

Model Name	Quality	Speed	Stability	Cost-Effectiveness	Vision	Thinking	Notes
claude-4-5	👑	🥈	🥈	🥈	👁️	🧠	Anthropic's latest high-quality flagship coding model
gemini-3.1-pro-preview	👑	🥈	🥇	🥉	👁️	🧠	Google's latest generation flagship model, excellent performance in the programming field, extremely strict logic ⚠️ Preview model
gemini-2.5-pro	🥇	🥇	🥇	🥈	👁️	🧠	Classic flagship model, stable programming ability, supports adaptive thinking ⚠️ Expected to be discontinued on June 17, 2026

After the generation model generates modification suggestions, we also need to use the Plugin Code Application Model Group (PLUGIN_APPLY_MODEL_GROUP) to apply the modification suggestions in the current plugin editor. It is recommended to use models with strong prompt compliance and fast generation speed. Below is the list of recommended models:

Model Name	Quality	Speed	Stability	Cost-Effectiveness	Vision	Thinking	Notes
gemini-3-flash-preview	🥇	👑	🥇	👑	👁️	❌	Recommended fast logic application model
gemini-2.5-flash	🥈	👑	🥇	🥈	👁️	❌	⚠️ Expected to be discontinued on June 17, 2026

Built-in Plugins

Emoticon Pack Plugin

The emoticon pack plugin needs to use a Vector Embedding Model to provide emoticon search capability. It is strongly recommended to use the text-embedding-v3 model:

Model Name	Quality	Speed	Stability	Cost-Effectiveness	Vision	Dimensions	Notes
text-embedding-v3	👑	👑	👑	👑	❌	1024	Very cheap and efficient text embedding model provided by Alibaba Cloud
multimodal-embedding-v1	👑	🥇	👑	👑	✅	1024	Multimodal embedding model provided by Alibaba Cloud, but with many input restrictions, only recommended for special use

Drawing (Learn to Draw)

The drawing plugin supports OpenAI standard drawing API (such as DALL-E 3) and any OpenAI chat completion API that supports conversation-generated images. Below is the list of recommended models:

Model Name	Quality	Speed	Stability	Cost-Effectiveness	Image-to-Image	Format	Notes
gemini-3.1-flash-image-preview	👑	🥇	🥇	🥇	✅	Chat mode	Gemini 3.1 drawing model, with extremely high understanding and visual quality
gemini-3-pro-image-preview	👑	🥇	🥈	🥉	✅	Chat mode	Gemini 3 flagship drawing model, rich in details
sora_image	🥇	⚪	🥇	🥈	✅	Chat mode	Consistent with ChatGPT official website 4o drawing, good logic compliance but slow
Kolors	🥈	👑	👑	🥇	✅	Image generation mode	Classic domestic drawing model, suitable for CG style tasks

Notes

1. Model performance may change over time with updates
2. Price information is for reference only, actual prices are subject to official quotations
3. It is recommended to regularly evaluate model selection based on actual usage
4. Experimental Models (exp/preview): These models are experimental and may be updated or closed at any time. It is recommended that:
   • Regularly follow Google Gemini API Version Notes for the latest updates
   • Prepare backup solutions when using in production environments
   • Prioritize using stable version (GA) models
   • Some preview models will automatically redirect to stable versions. It is recommended to directly use stable version model names to avoid delays caused by redirection
5. Model Redirection: Some discontinued preview models will automatically redirect to corresponding stable versions, for example:
   • gemini-3-pro-preview → gemini-3.1-pro-preview
   • gemini-2.5-flash-image-preview → gemini-3.1-flash-image-preview
   • gemini-2.5-pro-preview-06-05 → gemini-2.5-pro

Important Note

When using any generative artificial intelligence service, be sure to comply with relevant terms of service and laws and regulations