📑 Table of Contents

  1. Why It Exists

  2. What It Can Actually Do

  3. Before You Start

  4. Installation

  5. Setup Step by Step

  6. Prompts to Use It

  7. Daily Workflow

  8. Tips

  9. Quick Reference

🎯 Why It Exists

You ask one model a hard research question. It sounds confident. It misses a source, contradicts itself across paragraphs, and you have no way to know what it skipped without running the same prompt through three other models and reading all four by hand.

OpenRouter tested this on the DRACO deep research benchmark: 100 tasks across 10 domains, from law and medicine to finance and product comparison. A single frontier model topped out around 60%. But when they ran a panel of models in parallel and let a judge model fuse the results, scores jumped past every solo model.

The deeper finding: diversity beats raw horsepower. A panel of three budget models landed within 1% of Claude Fable 5 at roughly half the cost. Fusion exists so you get that panel effect from one API call instead of orchestrating it yourself.

⚡ What It Can Actually Do

You say

What happens

Send a prompt to openrouter/fusion

Your prompt fans out to a panel of models in parallel, each with web search and web fetch enabled

Add {"type": "openrouter:fusion"} to your tools array

The calling model decides on its own when a prompt needs deliberation, then invokes the panel

Define a custom analysis_models list

You pick exactly which models sit on the panel

Set a custom model (judge)

You choose which model synthesizes the final answer

Ask a complex research question

A judge reads every panel response and extracts consensus, contradictions, partial coverage, unique insights, and blind spots, then a synthesizer writes the final grounded answer

❝

⚠️ Fusion only triggers on prompts that genuinely need deliberation when used as a tool. Short tactical prompts will not fire it. Force it with the openrouter/fusion model alias if you want it on every request.

✅ Before You Start

  • An OpenRouter account → openrouter.ai

  • Credits added to your account (Fusion bills the underlying model usage)

  • An API key from your OpenRouter dashboard

  • Any OpenAI compatible HTTP client (cURL, the OpenAI SDK, or fetch)

No new SDK, no install of a Fusion specific package. If you already call OpenRouter, you already have everything.

💻 Installation

Fusion is a server side feature. There is nothing to install locally beyond an HTTP client you already use. The steps below cover the optional SDK setup per OS.

🍎 Mac

bash

# Option A: cURL is preinstalled, nothing to do
# Option B: use the OpenAI SDK (Fusion is OpenAI compatible)
pip3 install openai

🪟 Windows

powershell

# Option A: cURL ships with Windows 10/11, nothing to do
# Option B: use the OpenAI SDK
pip install openai

🐧 Linux

bash

pip install openai
❝

⚠️ You do NOT install Fusion itself. It runs entirely on OpenRouter's servers. You only need a way to send an HTTP POST request.

🛠 Setup Step by Step

  1. Go to openrouter.ai and sign in.

  2. Open your dashboard and add credits to your account.

  3. Create an API key and copy it somewhere safe.

  4. Set the key as an environment variable so it stays out of your code:

bash

   export OPENROUTER_API_KEY="your_key_here"

  1. Send your first Fusion call. Use the direct model slug openrouter/fusion:

bash

   curl https://openrouter.ai/api/v1/chat/completions \
     -H "Authorization: Bearer $OPENROUTER_API_KEY" \
     -H "Content-Type: application/json" \
     -d '{
       "model": "openrouter/fusion",
       "messages": [
         { "role": "user", "content": "What are the strongest arguments for and against carbon taxes?" }
       ]
     }'

  1. Read the response. The final answer comes back exactly like a single model response, grounded in the panel analysis.

  2. (Optional) Prefer to let the model decide when to deliberate? Skip the slug and attach Fusion as a tool instead:

json

   {
     "model": "anthropic/claude-opus-4.8",
     "messages": [{ "role": "user", "content": "..." }],
     "tools": [{ "type": "openrouter:fusion" }]
   }

  1. (Optional) Build a custom panel. Pick your own participant models and judge:

json

   {
     "model": "openrouter/fusion",
     "messages": [{ "role": "user", "content": "..." }],
     "plugins": [{
       "id": "fusion",
       "model": "google/gemini-3-flash-preview",
       "analysis_models": [
         "google/gemini-3-flash-preview",
         "moonshotai/kimi-k2.6",
         "deepseek/deepseek-v4-pro"
       ]
     }]
   }

💬 Prompts to Use It

Deep research

Survey the current evidence on [topic]. Identify where credible
sources disagree, and flag any claims that are contested.

Compare and contrast

Compare [option A] and [option B] across cost, performance, and
risk. Where would each one win, and where do experts disagree?

Expert critique

Here is a plan: [paste plan]. Have the panel stress test it.
Surface blind spots and contradictions before I commit to it.

Multi domain question

Answer this question that spans both [domain 1] and [domain 2]:
[question]. Ground every claim in a source.

🔁 Daily Workflow

How a real person uses this day to day:

  1. Default to solo models for fast, tactical prompts. Fusion is overkill for "rewrite this sentence."

  2. Reach for Fusion when being wrong is expensive. Legal research, financial analysis, medical questions, architecture decisions.

  3. Start with the default panel by calling openrouter/fusion directly. Only build a custom panel once you know which models you trust for your domain.

  4. For cost sensitive work, build a budget panel. Gemini 3 Flash plus Kimi K2.6 plus DeepSeek V4 Pro landed within 1% of Fable 5 at roughly half the cost on the benchmark.

  5. Let the agent decide in production. Attach Fusion as a tool so short prompts stay cheap and only the hard ones trigger the panel.

💡 Tips

  • Diversity is the lever, not size. The benchmark showed a panel of different budget models can rival a single frontier model. Mix architectures, do not just stack expensive ones.

  • Synthesis alone adds lift. OpenRouter fused Opus 4.8 with itself and still gained 6.7 points over solo Opus 4.8. Running the same prompt twice produces different reasoning paths the judge can reconcile. Even a same model panel helps on hard tasks.

  • Use exclude lists when benchmarking. If your panel has web search on, it can stumble onto answer keys or rubrics. Pass excluded_domains to keep evals honest.

  • Pick your judge deliberately. The judge model writes the final answer. A weak judge can waste a strong panel.

  • Tool mode keeps costs sane. Attaching Fusion as a tool means it only fires on prompts that need it, not every call.

❝

⚠️ Server tools are in beta. The API and behavior may change. Check the Fusion docs before relying on it in production.

📌 Quick Reference

json

// Force Fusion on every request
{ "model": "openrouter/fusion", "messages": [{ "role": "user", "content": "..." }] }

// Let the model decide when to use it
{
  "model": "anthropic/claude-opus-4.8",
  "messages": [{ "role": "user", "content": "..." }],
  "tools": [{ "type": "openrouter:fusion" }]
}

// Custom budget panel (within 1% of Fable 5 at roughly half the cost on DRACO)
{
  "model": "openrouter/fusion",
  "messages": [{ "role": "user", "content": "..." }],
  "plugins": [{
    "id": "fusion",
    "model": "google/gemini-3-flash-preview",
    "analysis_models": [
      "google/gemini-3-flash-preview",
      "moonshotai/kimi-k2.6",
      "deepseek/deepseek-v4-pro"
    ]
  }]
}

Endpoint: https://openrouter.ai/api/v1/chat/completions Try it in a chatroom: openrouter.ai/fusion API docs: openrouter.ai/docs/guides/features/server-tools/fusion

By The AI Leverage - Learn and master AI daily

Keep Reading