June 15 Message, ModelAggregation Platform OpenRouter The release of Fusion allows users to send the same hint to a set of models in parallel, as is the case with individual models, and then to analyse consensus, contradictions, blind areas and unique information by the Judice model, and eventually synthesize the answers。
Officially, 100 missions were tested on the DRACO in-depth study benchmark. The budget panel consisting of Gemini 3 Flash, Kimi K2.6 and DeepSeek V4 Pro scored 64.7%, close to the Fable 5 model, with a cost of about half。
Fusion supports direct calls to openrouter-fusion, or can customise participation and judge models. OpenRouter claims that the whole pipe is implemented at the service end, that the participating model is defaulted to have web-search and web-fetch capabilities and is suitable for research issues requiring cross-certification, blinding and combined output。