OpenRouter Launches Fusion: Multi-Model Consensus System

OpenRouter has released Fusion, a new routing system that runs user prompts through multiple AI models simultaneously before synthesizing their responses into a single answer.

How Fusion Works

Fusion processes each prompt through a panel of expert models running in parallel, with web search and web fetch capabilities enabled. A separate judge model then analyzes all responses to identify:

• Consensus points across models
• Contradictions between responses
• Partial coverage gaps
• Unique insights from individual models
• Blind spots in the collective analysis

The judge model uses this structured analysis to generate the final answer.

Configuration and Pricing

By default, Fusion uses OpenRouter's Quality preset for the expert panel. Users can switch to a Budget preset for lower-cost models or completely override both the panel and judge model using the fusion plugin's analysis_models and model fields.

Pricing is calculated as the sum of all underlying model completions. Since Fusion runs every panel member plus a judge call, costs are significantly higher than single-model requests. Users can view which models ran for each request in OpenRouter's Activity section.

Target Use Cases

According to OpenRouter, Fusion is designed for scenarios where "the cost of being wrong outweighs a few extra completions." Recommended applications include:

• Research requiring multiple perspectives
• Expert critique and analysis
• High-stakes decisions where accuracy is critical

The default panel includes six models from the Quality preset, though OpenRouter has not disclosed specific model names or which model serves as the judge.

Technical Details

Fusion is accessed through OpenRouter's routing system at openrouter/fusion. The service integrates with OpenRouter's existing web search infrastructure and can be customized through plugin parameters.

OpenRouter also offers an Auto Router as an alternative routing method, though details on how it differs from Fusion were not provided.

What This Means

Fusion represents a shift from single-model inference to ensemble approaches for high-stakes queries. By running multiple models in parallel and synthesizing their outputs, it trades cost and latency for potentially higher accuracy through consensus. However, without disclosed benchmark comparisons or specific model configurations, the practical accuracy gains remain unclear. The pricing model—sum of all completions—makes this approach expensive for routine queries, positioning it strictly as a tool for critical decisions where multiple expert opinions justify the cost premium.