Model Selection

Choosing the Right Model

For Legal Document Analysis

• Best: anthropic/claude-sonnet-4.5 - Excellent at long documents, nuanced reasoning
• Fast: anthropic/claude-4.5-haiku - Quick summaries, fast turnaround
• Deep Analysis: anthropic/claude-opus-4.1 - Most thorough, best for complex cases

For Coding/Structured Output

• Best: openai/gpt-5-codex - Optimized for code generation
• Fast: deepseek/deepseek-v3.1 - Great reasoning, very affordable

For Cost-Sensitive Tasks

• Ultra-cheap: alibaba/qwen-3-32b - $0.0001/1M input tokens
• Balanced: deepseek/deepseek-v3.1 - $0.0003/1M tokens (excellent quality)
• Fast & Cheap: xai/grok-4-fast-non-reasoning - $0.0002/1M tokens

For Medical Records

• Best: anthropic/claude-opus-4.1 - Understands medical terminology deeply
• Vision: anthropic/claude-sonnet-4.5 - Can read charts, scans, handwriting
• Budget: google/gemini-2.5-flash - Good reasoning at lower cost

For Reasoning Tasks

• Best: openai/gpt-5-pro - Deepest thinking, multi-hour reasoning
• Fast: openai/o4-mini - Quick reasoning for math/logic
• Open: deepseek/deepseek-r1 - Transparent reasoning traces

Understanding Response Metadata

Provider Metadata

The API routes requests through multiple providers for reliability:

{
  "provider_metadata": {
    "gateway": {
      "routing": {
        "originalModelId": "anthropic/claude-4.5-haiku",
        "resolvedProvider": "anthropic",
        "fallbacksAvailable": ["vertexAnthropic", "bedrock"],
        "finalProvider": "anthropic",
        "attempts": [
          {
            "provider": "anthropic",
            "credentialType": "byok",
            "success": true,
            "statusCode": 200
          }
        ]
      },
      "cost": "0.0000336",
      "marketCost": "0.0000336"
    }
  }
}

• originalModelId: The model you requested
• resolvedProvider: Which provider fulfilled the request
• fallbacksAvailable: Backup providers if primary fails
• credentialType: byok (bring your own key) or system (CaseMark's keys)
• attempts: Shows retry history if there were failures

Finish Reasons

• stop: Model naturally completed the response
• length: Hit max_tokens limit (response may be incomplete)
• content_filter: Response blocked by safety filters
• tool_calls: Model wants to call a function (for tool use)

Model Selection Cheat Sheet