Skip to main content
Use Fusion when one answer benefits from several model perspectives. In the Playground, Fusion configures multiple named ai-stats:advisor tools for analysis and an optional judge model for synthesis guidance. Fusion is a convenience workflow built on Advisor. There is no separate ai-stats:fusion API tool type; API clients can create the same behavior by adding multiple named Advisor tools.

​
Playground setup

In the Playground server-tools menu, Fusion lets you choose:
  • analysis models that independently inspect the request
  • an optional judge model that compares the analysis outputs
  • a max tool-call limit for the Fusion advisor tools
The main model sees those named advisors during the same tool loop and can use their outputs before writing the final response.

​
Equivalent API shape

{
  "tools": [
    {
      "type": "ai-stats:advisor",
      "parameters": {
        "name": "fusion_analysis_1",
        "model": "anthropic/claude-opus-4-8",
        "instructions": "Analyze the user's request independently. Return concise findings, assumptions, caveats, and the answer direction you recommend for synthesis.",
        "max_uses": 1,
        "max_completion_tokens": 1400
      }
    },
    {
      "type": "ai-stats:advisor",
      "parameters": {
        "name": "fusion_analysis_2",
        "model": "openai/gpt-5",
        "instructions": "Analyze the user's request independently. Return concise findings, assumptions, caveats, and the answer direction you recommend for synthesis.",
        "max_uses": 1,
        "max_completion_tokens": 1400
      }
    },
    {
      "type": "ai-stats:advisor",
      "parameters": {
        "name": "fusion_judge",
        "model": "google/gemini-pro-latest",
        "instructions": "Compare the fusion analysis outputs, identify the strongest answer, and return concise guidance for the final response.",
        "max_uses": 1,
        "max_completion_tokens": 1400
      }
    }
  ]
}

​
When to use Fusion

Use Fusion for:
  • planning decisions where a second or third model may catch missing assumptions
  • research synthesis where models may weigh evidence differently
  • code or architecture reviews where you want independent critique
  • high-value responses where a little extra latency is acceptable
Use a single Advisor instead when you only need one review pass.

​
Usage and pricing

Fusion uses Advisor calls under the hood. Each analysis or judge call increments: 
{
  "usage": {
    "server_tool_use": {
      "advisor_requests": 3
    }
  }
}
The advisor models’ tokens are included in total usage and can be priced at each selected advisor model’s rates. Server-tool pricing can also use server_tool_advisor_requests.
Last modified on July 2, 2026