openrouter（OpenRouter 模型网关）

OpenRouter Skill

Comprehensive assistance with OpenRouter API development, providing unified access to hundreds of AI models through a single endpoint with intelligent routing, automatic fallbacks, and standardized interfaces.

When to Use This Skill

This skill should be triggered when:

• Making API calls to multiple AI model providers through a unified interface
• Implementing model fallback strategies or auto-routing
• Working with OpenAI-compatible SDKs but targeting multiple providers
• Configuring advanced sampling parameters (temperature, top_p, penalties)
• Setting up streaming responses or structured JSON outputs
• Comparing costs across different AI models
• Building applications that need automatic provider failover
• Implementing function/tool calling across different models
• Questions about OpenRouter-specific features (routing, fallbacks, zero completion insurance)

Quick Reference

Basic Chat Completion (Python)

from openai import OpenAI

client = OpenAI(
  base_url="https://openrouter.ai/api/v1",
  api_key="<OPENROUTER_API_KEY>",
)

completion = client.chat.completions.create(
  model="openai/gpt-4o",
  messages=[{"role": "user", "content": "What is the meaning of life?"}]
)
print(completion.choices[0].message.content)

Basic Chat Completion (JavaScript/TypeScript)

import OpenAI from 'openai';

const openai = new OpenAI({
  baseURL: 'https://openrouter.ai/api/v1',
  apiKey: '<OPENROUTER_API_KEY>',
});

const completion = await openai.chat.completions.create({
  model: 'openai/gpt-4o',
  messages: [{"role": 'user', "content": 'What is the meaning of life?'}],
});
console.log(completion.choices[0].message);

cURL Request

curl https://openrouter.ai/api/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  -d '{
    "model": "openai/gpt-4o",
    "messages": [{"role": "user", "content": "What is the meaning of life?"}]
  }'

Model Fallback Configuration (Python)

completion = client.chat.completions.create(
    model="openai/gpt-4o",
    extra_body={
        "models": ["anthropic/claude-3.5-sonnet", "gryphe/mythomax-l2-13b"],
    },
    messages=[{"role": "user", "content": "Your prompt here"}]
)

Model Fallback Configuration (TypeScript)

const completion = await client.chat.completions.create({
    model: 'openai/gpt-4o',
    models: ['anthropic/claude-3.5-sonnet', 'gryphe/mythomax-l2-13b'],
    messages: [{ role: 'user', content: 'Your prompt here' }],
});

Auto Router (Dynamic Model Selection)

completion = client.chat.completions.create(
    model="openrouter/auto",  # Automatically selects best model for the prompt
    messages=[{"role": "user", "content": "Your prompt here"}]
)

Advanced Parameters Example

completion = client.chat.completions.create(
    model="openai/gpt-4o",
    messages=[{"role": "user", "content": "Write a creative story"}],
    temperature=0.8,           # Higher for creativity (0.0-2.0)
    max_tokens=500,            # Limit response length
    top_p=0.9,                 # Nucleus sampling (0.0-1.0)
    frequency_penalty=0.5,     # Reduce repetition (-2.0-2.0)
    presence_penalty=0.3       # Encourage topic diversity (-2.0-2.0)
)

Streaming Response

stream = client.chat.completions.create(
    model="openai/gpt-4o",
    messages=[{"role": "user", "content": "Tell me a story"}],
    stream=True
)

for chunk in stream:
    if chunk.choices[0].delta.content:
        print(chunk.choices[0].delta.content, end='')

JSON Mode (Structured Output)

completion = client.chat.completions.create(
    model="openai/gpt-4o",
    messages=[{
        "role": "user",
        "content": "Extract person's name, age, and city from: John is 30 and lives in NYC"
    }],
    response_format={"type": "json_object"}
)

Deterministic Output with Seed

completion = client.chat.completions.create(
    model="openai/gpt-4o",
    messages=[{"role": "user", "content": "Generate a random number"}],
    seed=42,            # Same seed = same output (when supported)
    temperature=0.0     # Deterministic sampling
)

Key Concepts

Model Routing

OpenRouter provides intelligent routing capabilities:

• Auto Router (openrouter/auto): Automatically selects the best model based on your prompt using NotDiamond
• Fallback Models: Specify multiple models that automatically retry if primary fails
• Provider Routing: Automatically routes across providers for reliability

Authentication

• Uses Bearer token authentication with API keys
• API keys can be managed programmatically
• Compatible with OpenAI SDK authentication patterns

Model Naming Convention

Models use the format provider/model-name:

• openai/gpt-4o - OpenAI's GPT-4 Optimized
• anthropic/claude-3.5-sonnet - Anthropic's Claude 3.5 Sonnet
• google/gemini-2.0-flash-exp:free - Google's free Gemini model
• openrouter/auto - Auto-routing system

Sampling Parameters

Temperature (0.0-2.0, default: 1.0)

• Lower = more predictable, focused responses
• Higher = more creative, diverse responses
• Use low (0.0-0.3) for factual tasks, high (0.8-1.5) for creative work

Top P (0.0-1.0, default: 1.0)

• Limits choices to percentage of likely tokens
• Dynamic filtering of improbable options
• Balance between consistency and variety

Frequency/Presence Penalties (-2.0-2.0, default: 0.0)

• Frequency: Discourages repeating tokens proportional to use
• Presence: Simpler penalty not scaled by count
• Positive values reduce repetition, negative encourage reuse

Max Tokens (integer)

• Sets maximum response length
• Cannot exceed context length minus prompt length
• Use to control costs and enforce concise replies

Response Formats

• Standard JSON: Default chat completion format
• Streaming: Server-Sent Events (SSE) with stream: true
• JSON Mode: Guaranteed valid JSON with response_format: {"type": "json_object"}
• Structured Outputs: Schema-validated JSON responses

Advanced Features

• Tool/Function Calling: Connect models to external APIs
• Multimodal Inputs: Support for images, PDFs, audio
• Prompt Caching: Reduce costs for repeated prompts
• Web Search Integration: Enhanced responses with web data
• Zero Completion Insurance: Protection against failed responses
• Logprobs: Access token probabilities for confidence analysis

Reference Files

This skill includes comprehensive documentation in references/:

• llms-full.md - Complete list of available models with metadata
• llms-small.md - Curated subset of popular models
• llms.md - Standard model listings

Use view to read specific reference files when detailed model information is needed.

Working with This Skill

For Beginners

1. Start with basic chat completion examples (Python/JavaScript/cURL above)
2. Use the standard OpenAI SDK for easy integration
3. Try simple model names like openai/gpt-4o or anthropic/claude-3.5-sonnet
4. Keep parameters simple initially (just model and messages)

For Intermediate Users

1. Implement model fallback arrays for reliability
2. Experiment with sampling parameters (temperature, top_p)
3. Use streaming for better UX in conversational apps
4. Try openrouter/auto for automatic model selection
5. Implement JSON mode for structured data extraction

For Advanced Users

1. Fine-tune multiple sampling parameters together
2. Implement custom routing logic with fallback chains
3. Use logprobs for confidence scoring
4. Leverage tool/function calling capabilities
5. Optimize costs by selecting appropriate models per task
6. Implement prompt caching strategies
7. Use seed parameter for reproducible testing

Common Patterns

Error Handling with Fallbacks

try:
    completion = client.chat.completions.create(
        model="openai/gpt-4o",
        extra_body={
            "models": [
                "anthropic/claude-3.5-sonnet",
                "google/gemini-2.0-flash-exp:free"
            ]
        },
        messages=[{"role": "user", "content": "Your prompt"}]
    )
except Exception as e:
    print(f"All models failed: {e}")

Cost-Optimized Routing

# Use cheaper models for simple tasks
simple_completion = client.chat.completions.create(
    model="google/gemini-2.0-flash-exp:free",
    messages=[{"role": "user", "content": "Simple question"}]
)

# Use premium models for complex tasks
complex_completion = client.chat.completions.create(
    model="openai/o1",
    messages=[{"role": "user", "content": "Complex reasoning task"}]
)

Context-Aware Temperature

# Low temperature for factual responses
factual = client.chat.completions.create(
    model="openai/gpt-4o",
    temperature=0.2,
    messages=[{"role": "user", "content": "What is the capital of France?"}]
)

# High temperature for creative content
creative = client.chat.completions.create(
    model="openai/gpt-4o",
    temperature=1.2,
    messages=[{"role": "user", "content": "Write a unique story opening"}]
)

Resources

Official Documentation

• API Reference: https://openrouter.ai/docs/api-reference/overview
• Quickstart Guide: https://openrouter.ai/docs/quickstart
• Model List: https://openrouter.ai/docs/models
• Parameters Guide: https://openrouter.ai/docs/api-reference/parameters

Key Endpoints

• Chat Completions: POST https://openrouter.ai/api/v1/chat/completions
• List Models: GET https://openrouter.ai/api/v1/models
• Generation Info: GET https://openrouter.ai/api/v1/generation

Notes

• OpenRouter normalizes API schemas across all providers
• Uses OpenAI-compatible API format for easy migration
• Automatic provider fallback if models are rate-limited or down
• Pricing based on actual model used (important for fallbacks)
• Response includes metadata about which model processed the request
• All models support streaming via Server-Sent Events
• Compatible with popular frameworks (LangChain, Vercel AI SDK, etc.)

Best Practices

1. Always implement fallbacks for production applications
2. Use appropriate temperature based on task type (low for factual, high for creative)
3. Set max_tokens to control costs and response length
4. Enable streaming for better user experience in chat applications
5. Use JSON mode when you need guaranteed structured output
6. Test with seed parameter for reproducible results during development
7. Monitor costs by selecting appropriate models per task
8. Use auto-routing when unsure which model performs best
9. Implement proper error handling for rate limits and failures
10. Cache prompts for repeated requests to reduce costs