OpenRouter has introduced Fusion, an API that sends a prompt to several language models in parallel and then uses a separate judge model to compare and consolidate their answers. The goal is to remove the need for developers to act as manual routers — querying models one by one and selecting the best response themselves.
How it works
- You send a prompt to Fusion.
- Fusion dispatches the task in parallel to a panel of models.
- Models are given web search and web fetch access to inform their responses.
- A judge model compares the returned answers and produces the final result.
Test results and concrete numbers
OpenRouter’s internal tests indicate Fusion can improve performance when models are combined. Key results include:
- A Fusion panel combining Fable 5 and GPT-5.5 scored 69.0% on the DRACO benchmark, outperforming every single model in the test.
- Fable 5 alone scored 65.3%, and GPT-5.5 alone scored 60.0%.
- A budget panel consisting of Gemini 3 Flash, Kimi K2.6, and DeepSeek V4 Pro scored 64.7%; this beat GPT-5.5 and Claude Opus 4.8 individually and came within roughly one point of Fable 5, while costing about half as much.
These results show that pairing frontier models can yield gains beyond any one model’s standalone ceiling, and that cost-effective ensembles can approach top-performing single models.
Why this matters
Fusion highlights that model diversity is becoming a product primitive: rather than choosing a single model and hoping it’s sufficient, builders can treat a set of models as a small, complementary research team. This approach can improve accuracy, robustness, and cost-performance trade-offs in production systems.
How to use it
Fusion is available through the regular OpenRouter API. You can call the openrouter/fusion endpoint directly or configure the Fusion plugin with your own analysis models and judge model. That lets teams customize which models are in the panel and how the judge evaluates outputs.
Summary
OpenRouter Fusion combines parallel model invocations, web-enabled context, and a judge model to return a single vetted answer, reducing the need for bespoke routing logic. DRACO benchmark results from OpenRouter indicate that intelligently combining models can outperform individual models and enable cheaper, competitive panel configurations.
