OpenRouter Fusion: When Two AIs Team Up and Beat Every Single Model

Links & Resources

• Fusion Chatroom
• Announcement Blog Post
• API Docs
• DRACO Benchmark Paper
• OpenRouter

Claude Fable 5 just lost to itself. Not to a bigger model. Not to a smarter model. To two AIs double-checking each other's work.

OpenRouter dropped Fusion on June 12, 2026 - a tool that fires your question at a panel of models simultaneously, then has a judge model read every answer and stitch the best parts into one final response. The result beats every single model that exists, including Fable 5.

The Headline Numbers

Tested on 100 deep research tasks from the DRACO benchmark (by Perplexity AI), which scores reasoning, tool usage, source quality, and factual accuracy across 10 domains:

Three things jump out of this table.

1. Fable 5 + GPT-5.5 Beat Everything

Fable 5 alone scored 65.3%. GPT-5.5 alone scored 60.0%. Fuse them together with Opus 4.8 as judge and the combo hits 69.0% - beating every individual model on the planet.

Neither model could reach that score on its own. The performance gain isn't from picking the better answer - it's from the synthesis step combining reasoning paths, source selections, and tool calls that neither model found alone.

2. Budget Models Beat Frontier Models at Half the Price

A team of three cheap models - Gemini 3 Flash, Kimi K2.6, and DeepSeek V4 Pro - scored 64.7%. That beats GPT-5.5 (60.0%) and Claude Opus 4.8 (58.8%) and comes within 1% of Fable 5 (65.3%).

The cost? Roughly half what you'd pay for a single frontier model call.

This is the real unlock. You don't need the most expensive model to get frontier-level answers. You need a team of budget models that check each other's work.

3. Same Model + Same Model = Smarter

This is the wildest result. OpenRouter fused Opus 4.8 with itself - same model, asked twice, same judge. The score jumped from 58.8% to 65.5% - a 6.7-point gain from asking the same AI twice.

Why? Because the same prompt produces different reasoning paths, different tool calls, different source selections each time. The synthesis step catches what one pass missed and resolves contradictions between the two runs.

How Fusion Works

The pipeline runs server-side in a single API call:

Step 1 - Parallel dispatch. Your prompt goes to every model in the panel simultaneously. Each model has access to web search, web fetch, and bash - identical tool sets across the board.

Step 2 - Structured analysis. The judge model reads every panel response and produces a structured breakdown: consensus points, contradictions, partial coverage, unique insights, and blind spots.

Step 3 - Synthesis. The judge writes the final answer grounded in that analysis - pulling the best reasoning from each model, resolving conflicts, and filling gaps none of them covered individually.

The whole thing is one API call. You send it like you would any model request.

4 Ways to Use Fusion

1. Chatroom (Easiest)

Go to openrouter.ai/fusion, pick a preset panel or build a custom one, and start chatting. No code.

2. Model Slug (One-Line Swap)

Replace your model string with openrouter/fusion and a default frontier panel is auto-injected:

{
  "model": "openrouter/fusion",
  "messages": [
    { "role": "user", "content": "Your question here" }
  ]
}

3. Server Tool (Model Decides When)

Add Fusion as a tool. Your base model handles routine tasks directly and calls Fusion only when a question is worth spending extra for multiple perspectives:

{
  "tools": [{ "type": "openrouter:fusion" }]
}

4. Plugin (Full Control)

Pick your judge model, pick your panel, and send it as a plugin:

{
  "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"
    ]
  }]
}

The DRACO Benchmark - Why This Test Matters

DRACO isn't a trivia quiz. It's 100 deep research tasks across 10 domains - academic research, finance, law, medicine, technology, UX design, general knowledge, needle-in-a-haystack retrieval, personalized assistance, and product comparison.

Each task is graded on ~39 weighted criteria across four categories:

• Factual Accuracy (~20 criteria) - verifiable claims the response must get right
• Breadth & Depth (~9 criteria) - synthesis quality, trade-off analysis, actionable guidance
• Presentation Quality (~6 criteria) - terminology, formatting, readability
• Citation Quality (~5 criteria) - primary source citations with working references

Criteria can carry negative weights. A model that confidently states wrong things gets punished. You can't game the score by being verbose - you win by being right.

Anti-Cheating

OpenRouter discovered that models with web search were finding the DRACO grading rubric online during evaluation. They solved this by excluding the hosting locations from search, ensuring no model could read its own test answers. All published results were produced after these exclusions were in place.

When to Use Fusion

Architecture decisions - when the cost of being wrong is high, getting multiple perspectives pays for itself.

Deep research - any question that needs synthesis across multiple sources, domains, or perspectives.

High-stakes analysis - legal research, medical questions, financial modeling where you need the answer checked before you act on it.

Budget frontier - when you need frontier-quality answers but can't justify frontier-model pricing. Three cheap models fused together get you 95% of the way there at 50% of the cost.

When NOT to Use Fusion

Routine coding - use your coding model directly. Fusion adds latency (2-3x longer than a standard call) that isn't worth it for straightforward tasks.

Simple questions - if one model can answer it in a sentence, don't invoke a panel.

Long-horizon agentic tasks - DRACO doesn't test these, and Fable 5 is still built for that use case.

The Takeaway

The best AI answer isn't one genius model thinking alone. It's a team double-checking its work. The same principle that makes human teams outperform individuals applies to AI - diverse perspectives, independent reasoning, structured synthesis.

And the math keeps getting better: as budget models improve, the floor of what Fusion can achieve rises with them.

Follow AI Adventure YT for more AI breakdowns and guides.