How Fusion Works

OpenRouter Fusion uses a three-phase pipeline. It's not ensemble blending or majority voting, it's structured comparison with a judge that identifies where models agreed, disagreed, and what they each uniquely contributed.

Your request
  → Your model (reads prompt, invokes openrouter:fusion tool)
    → Panel [2-8 models run in PARALLEL, each with web_search + web_fetch]
      → Judge model [receives all panel responses → structured JSON]
        → Your model [reads JSON analysis → writes final answer]

The judge JSON includes:

• Consensus points- high-confidence, all/most models agreed
• Contradictions- models disagreed, and on what specifically
• Partial coverage- some models addressed, others missed
• Unique insights per model- what only that model contributed
• Blind spots- nothing any model addressed

Three Ways to Call It

Method 1: Model alias (simplest):

client.chat.completions.create(
    model="openrouter/fusion",   # auto-injects default Quality panel
    messages=[{"role": "user", "content": "..."}],
    tool_choice="required"       # force it -- don't let the model skip
)

Method 2 -- Tool call (model decides when to invoke):

client.chat.completions.create(
    model="anthropic/claude-opus-4-7",   # your existing model
    messages=[...],
    tools=[{"type": "openrouter:fusion"}]
)

Method 3 -- Custom panel via plugins (full control):

# Reuse models already in our routing table:
"plugins": [{
    "id": "fusion",
    "preset": "general-budget",
    "analysis_models": [
        "deepseek/deepseek-chat",              # our short-form model
        "meta-llama/llama-3.3-70b-instruct",   # our classify model
        "~google/gemini-flash-latest"           # cheap panel member
    ],
    "model": "anthropic/claude-opus-4-7"       # Opus as judge -- judge matters most
}]

Key Constraints

• Context window: 128K tokens
• Panel size: 1–8 models
• Recursive prevention: built-in depth header stops nested Fusion calls
• Beta status- API and behavior may change
• Web search + web fetch: automatically enabled for every panel member and judge

Performance: The DRACO Benchmark

Pricing: The Honest Math

Fusion bills as the sum of all underlying completions- no blended rate, no ensemble discount. Pay for every panel call plus the judge call.

The "half the price of Fable 5" marketing claim is technically accurate for the Budget-vs-Fable-5 comparison only. Compared to Sonnet or Opus, there's no cost advantage, just a quality argument.

Fusion vs. Our Adversarial Reviews

This is the most important comparison. They're solving different problems with opposite epistemic frames.

When to Use Fusion in Our Workflow

Use Fusion (Budget preset)

• Use /research skill, this is exactly what DRACO benchmarks. Replace Opus direct with Fusion Budget.
• Use auto-research- replace the manual 5–7 agent fan-out. Fusion consolidates in 1 API call with web search built in.
• Use /market-size and /challenge IdeaForge skills, adversarial multi-perspective analysis fits.
• Use /weekly-review synthesis, multi-source synthesis across the week.
• Use Hub deliverables with medical/wellness claims (SmartStrongAlive, BoneVoyageRevival), high-stakes, no easy ground truth.

Skip Fusion: Keep Existing Routing

• Skip All code generation → DeepSeek V3.1 (Fusion has no coding validation)
• Skip Triage/classify → Llama 70B (wrong shape entirely)
• Skip Adversarial code/plan review → keep existing Codex + Gemini + Grok pattern
• Skip /morning, /evening check-ins → single model, latency matters
• Skip High-volume batch inference → cost multiplier becomes painful at scale
• Skip Any prompt containing private data (BrainOS, PHI, prospect data, financial details)

Minimal Viable Integration

Three steps to add Fusion to our existing OpenRouter workflow without touching anything that's already working:

Step 1: Test Budget preset on the next auto-research task:

# Swap for research tasks in any existing OpenRouter script:
model = "openrouter/fusion"
tool_choice = "required"   # force invocation, don't let model skip it

Step 2: Custom panel reusing our models:

plugins = [{
    "id": "fusion",
    "preset": "general-budget",
    "analysis_models": [
        "deepseek/deepseek-chat",
        "meta-llama/llama-3.3-70b-instruct",
        "~google/gemini-flash-latest"
    ],
    "model": "anthropic/claude-opus-4-7"  # Opus as judge -- judge matters most
}]

Step 3: Wire a DEEP_RESEARCH flag in the skill router: When a skill step is tagged DEEP_RESEARCH: true, route to Fusion Budget. Everything else routes as today (Llama, DeepSeek, Opus direct).

Real-World Gotchas

• Cache buster: Fusion destroys prompt caching, each panel call is fresh. Don't use it on high-volume pipelines that rely on caching economics.
• Latency: 1–3 seconds overhead minimum (slowest panel member is the bottleneck). One real-world test found 7x slower end-to-end on Quality preset. Not for real-time UX.
• Judge identity is opaque: OpenRouter uses alias slugs (~anthropic/claude-opus-latest), not pinned model IDs. Specific models in each preset will change as new releases land.
• DRACO scores swing 10–25 points by judge model: The benchmark isn't as stable as the headline numbers suggest.
• Long-horizon tasks: Fusion showed no advantage for multi-step agentic chains. Single Fable 5 likely still wins there.
• Model-specific features are lost: Claude's extended thinking, OpenAI's code interpreter, Gemini's native multimodal, none carry through Fusion. You get text outputs only.
• Privacy footprint: Your prompt goes to every provider in the panel simultaneously. Treat with the same care as sending to a public API.

Contrarian View

75% of Fusion's quality gain comes from the synthesis step, not model diversity. That means you might get most of the benefit by prompting a single model to answer a question, then asking it to critique and revise its own output, at zero additional cost, zero latency overhead, and zero privacy risk. The self-critique loop is architecturally similar to what Fusion's judge does.

The Budget preset's DRACO gains are real, but they come from a benchmark OpenRouter controls, on tasks OpenRouter selected. Independent practitioner results on coding tasks have been flat or negative. Test on 10–20 real auto-research tasks before committing. Measure output quality subjectively. Adopt if it holds, don't adopt on benchmark marketing alone.