openrouter（OpenRouter 模型网关）

OpenRouter

Overview

OpenRouter provides a unified API to access hundreds of LLM models from different providers (OpenAI, Anthropic, Google, Meta, and more) with automatic routing, cost tracking, and performance monitoring. Use this skill to make API calls to any OpenRouter model, compare responses across models, track costs and latency, and optimize model selection.

Quick Start

To call an OpenRouter model:

1. Set OPENROUTER_API_KEY in your environment
2. Use the scripts/call_openrouter.sh script with --model and --prompt flags
3. Add --json flag for structured output

The script returns:

• Response time in seconds (wall-clock time)
• Cost in dollars (OpenRouter pricing)
• Full response content
• Token counts (prompt, completion, total)

Making API Calls

Basic Usage

The scripts/call_openrouter.sh script provides a flexible CLI interface:

# Basic call
bash scripts/call_openrouter.sh \
  --model "anthropic/claude-3.5-sonnet" \
  --prompt "Explain quantum computing" \
  --json

# With optional parameters
bash scripts/call_openrouter.sh \
  --model "openai/gpt-4o:nitro" \
  --prompt "Write a haiku" \
  --max-tokens 100 \
  --temperature 0.7 \
  --json

Command-Line Arguments

• --model (required): Model ID (e.g., "anthropic/claude-3.5-sonnet")
• --prompt (required): User prompt/question
• --system: Optional system message
• --max-tokens: Maximum tokens to generate
• --temperature: Temperature (0.0-2.0)
• --json: Output as JSON (default: human-readable)

Environment Variables

• OPENROUTER_API_KEY (required): Your API key
• OPENROUTER_REFERER (optional): HTTP referer for tracking (default: http://localhost)
• OPENROUTER_TITLE (optional): Title for tracking (default: Local Test)
• MODEL (optional): Override the default model

Reading the Output

The script outputs:

1. Response time in seconds (measured client-side)
2. Complete JSON response with:

- choices[0].message.content: The model's response - usage.prompt_tokens: Input token count - usage.completion_tokens: Output token count - usage.total_tokens: Total tokens used

Cost Calculation

To calculate costs:

1. Get the model's pricing from the models list (see references)
2. Calculate: (prompt_tokens × prompt_price) + (completion_tokens × completion_price)

Example: If a model costs $0.0000025/token for prompts and $0.000002/token for completions, and uses 14 prompt + 277 completion tokens:

• Cost = (14 × 0.0000025) + (277 × 0.000002) = $0.000035 + $0.000554 = $0.000589

Model Selection

Finding Models

Retrieve the full models list with pricing and capabilities:

curl https://openrouter.ai/api/v1/models -H "Authorization: Bearer $OPENROUTER_API_KEY" > models.json

The list is sorted by creation date (newest first), serving as a proxy for quality.

Important: The models list can be very large. Consider saving to a file and using grep/jq to filter by:

• Price range
• Context length
• Specific providers
• Capabilities (vision, function calling, etc.)

Model Naming Format

OpenRouter uses provider/model-name:

• anthropic/claude-3.5-sonnet
• openai/gpt-4o
• google/gemini-pro-1.5
• meta-llama/llama-3.1-405b-instruct

Speed and Feature Modifiers

:nitro - Use the fastest available provider for a model

anthropic/claude-3.5-sonnet:nitro

:online - Enable web search capabilities

openai/gpt-4o:online

Combine modifiers:

anthropic/claude-3.5-sonnet:nitro:online

Common Use Cases

Testing a Specific Model

Edit the script's PAYLOAD to use the desired model and messages:

{
  "model": "anthropic/claude-3.5-sonnet",
  "messages": [
    {"role": "user", "content": "Explain quantum computing in simple terms"}
  ]
}

Comparing Models

Run the script multiple times with different models and compare:

• Response quality
• Response time
• Token usage and cost

Finding the Cheapest/Fastest Model

1. Fetch the models list and save to file
2. Use jq or grep to filter by criteria
3. Test top candidates with the script
4. Compare performance vs. cost trade-offs

For speed: Try models with :nitro suffix For cost: Filter models.json by lowest pricing values

Accessing Provider Information (Non-API)

Opening Provider Pages with Query Parameters

While the OpenRouter API provides model information, provider-specific details like throughput, latency, and availability are only accessible via the web interface. You can programmatically open these pages with sorting parameters.

URL Structure

https://openrouter.ai/<model-slug>/providers?sort=<sorting-option>

Available Sorting Options:

• throughput - Sort by provider throughput (tokens/sec)
• price - Sort by cost
• latency - Sort by response latency

Example: Opening Provider Page Sorted by Throughput

For the model moonshotai/kimi-k2-0905:

https://openrouter.ai/moonshotai/kimi-k2-0905/providers?sort=throughput

Use Case: Finding the Fastest Provider

When you need to identify which provider offers the best throughput for a specific model:

1. Extract the model slug from the model ID (e.g., openai/gpt-4o → openai/gpt-4o)
2. Construct the URL: https://openrouter.ai/<model-slug>/providers?sort=throughput
3. Open the URL in a browser or use web automation tools
4. The page will display providers sorted by throughput (highest first)

Note: This information is not available through the API and requires web interface access. The :nitro modifier automatically routes to the fastest provider, but if you need to see provider-specific metrics, use the web interface with query parameters.

Workflow for Agent Tools

If you have browser automation capabilities:

• Use mcp__chrome-devtools__new_page or similar to open the provider page
• The ?sort=throughput parameter ensures the page loads pre-sorted
• Extract provider metrics from the rendered page

Accessing Model Rankings by Category

OpenRouter provides model rankings filtered by specific use cases and categories. These rankings show which models perform best for different tasks based on user ratings and token usage.

URL Structure

https://openrouter.ai/rankings?category=<category-value>#categories

Available Categories

Usage Notes

• Most categories use lowercase versions of their names (e.g., programming, science)
• The Marketing/Seo category uses marketing/seo with a slash
• The #categories anchor is optional but helps navigate to the categories section
• Rankings are not available through the API and require web interface access
• Each category shows models ranked by performance for that specific use case

Use Case: Finding the Best Model for a Specific Task

When you need to identify top-performing models for a particular domain:

1. Select the appropriate category from the table above
2. Construct the URL: https://openrouter.ai/rankings?category=<value>#categories
3. Open the URL in a browser or use web automation tools
4. The page displays models ranked by performance for that category

Example for programming tasks:

https://openrouter.ai/rankings?category=programming#categories

Workflow for Agent Tools

If you have browser automation capabilities:

• Use mcp__chrome-devtools__new_page to open the rankings page
• The ?category=<value> parameter loads the page with the selected category
• Verify the category dropdown shows the expected category name
• Extract model rankings and performance data from the rendered page

Resources

scripts/call_openrouter.sh

Bash script that makes an API call to OpenRouter and returns timing, cost, and full response. Uses curl and jq for simple, dependency-free execution.

Requirements: jq (for JSON parsing)

Usage:

bash call_openrouter.sh --model "anthropic/claude-3.5-sonnet" --prompt "Your question" --json

references/models_and_features.md

Detailed reference on:

• How to fetch and filter the models list
• Model naming conventions
• Speed (:nitro) and web search (:online) modifiers
• Cost calculation from usage data