What is Fusion API?

What is Fusion API?

Fusion API is a multi-model layer for language models. Instead of trusting a single model to think, search, and answer on its own, Fusion sends your prompt to a panel of strong models, has a judge model review every answer, and then synthesizes a single final result.

The idea is simple: a committee of models, coordinated well, produces more reliable answers than any one model working alone — often at a fraction of the cost of a single frontier model.

How Fusion works

A Fusion request runs in three stages:

1. Parallel research. Your prompt goes to several models at once. Each one researches and answers independently, with the same tool access — so you get genuinely different reasoning paths, not copies of the same answer.
2. Cross-review. A judge model reads every answer and maps where they agree, where they conflict, what they missed, and which claims look risky or wrong.
3. Synthesis. A lead model merges the reviewed findings into one final answer — keeping the consensus, resolving the conflicts, and dropping the weak claims.

The result is sharper than any single model in the panel, because the mistakes one model makes are usually caught by the others.

Why does a committee beat one model?

A single model commits to one line of reasoning. It might call one set of tools, read one set of sources, and fixate on one framing. Run several models in parallel and you get multiple independent attempts — and when you compare them, gaps and errors surface that a single pass would never reveal.

Interestingly, most of the gain comes from the review-and-synthesis step itself, not just from using different models. Even running the same model twice and synthesizing the two answers measurably improves quality, because the model can take different reasoning paths on each pass.

Is it actually better? The benchmark

On DRACO — a deep-research benchmark of 100 tasks across 10 domains (academic, financial, legal, medical, technical, and more), graded on accuracy, completeness, and citation reliability — a budget Fusion panel scores within roughly a point of a frontier model while costing about half as much. Solo frontier models, run alone, score several points lower than Fusion stacks built on cheaper models.

See the numbers and watch the flow on the live demo.

How to call Fusion via the API

Fusion is a routing layer on top of OpenRouter, so you call it like any other model — just point the model field at the openrouter/fusion alias:

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": "Postgres or SQLite for a serverless SaaS in 2026?" }
    ]
  }'

Behind that single call, a panel of up to 8 models answers in parallel (with web search and fetch), a judge model produces a structured review — consensus, contradictions, partial coverage, unique insights, and blind spots — and a final model writes the answer from it.

Presets and custom panels. OpenRouter ships two curated presets: general-high (the strongest panel) and general-budget (a cheaper panel with a frontier judge). You can also build your own panel with the analysis_models parameter (1–8 models) and pick a separate judge model. You don't have to touch JSON to experiment — set the panel, judge, and budget/quality preset right in our Playground, or estimate the cost first with the calculator.

Pricing. A Fusion call bills as the sum of the underlying completions — if the panel runs four models, you pay for those four plus the judge and synthesis. Even so, a budget panel reaches near-frontier quality at roughly half the cost of a premium single model.

When should you use Fusion?

Fusion shines on tasks where being right matters more than being fast:

• Deep research and analysis
• Technical and architectural decisions
• Legal, medical, and financial questions (where a wrong claim is costly)
• Anything you'd normally double-check across multiple sources

For trivial or latency-critical calls, a single fast model is still the better tool. Fusion is for the questions worth getting right.

How do I try it?

Run the interactive demo to watch the panel → judge → synthesis flow on real questions — free, right in your browser, no signup required.