OpenRouter Launches Fusion Composite Model to Outperform Single Models

According to Beating Monitor, OpenRouter has launched the Fusion composite model interface, which enables parallel distribution of user prompts to multiple large language models, followed by final answer generation through a judge model and synthesis model. In Perplexity AI’s DRACO deep research benchmark, various model combinations based on Fusion demonstrated superior performance over traditional single-model setups. Across all configurations, the panel combining Fable 5 and GPT-5.5, synthesized by Opus 4.8, achieved the highest score of 69.0%, significantly outperforming Fable 5’s standalone score of 65.3%. The tests indicate that the diversity arising from differences in underlying training and logic across models from different vendors enhances perspective complementarity in complex tasks. Even self-comparison within the same model shows notable improvement: when Opus 4.8 was paired with itself in a dual-panel setup for self-synthesis, its score rose from 58.8% (standalone) to 65.5%. For cost-efficient needs, a low-cost panel composed of Gemini 3 Flash, Kimi K2.6, and DeepSeek V4 Pro achieved a synthesized score of 64.7%, narrowing the gap with Fable 5 to just 0.6 percentage points while halving invocation costs. The benchmark comprises 100 complex research tasks across 10 dimensions. To prevent models from accidentally accessing online evaluation standards during retrieval—thereby distorting scores—OpenRouter configured a server-side filter list of restricted domains. Fusion is now available to API users, with the default identifier openrouter/fusion; users may also customize their synthesis model panels via the web interface.