**[OpenRouter Launches Fusion Composite Model: Multi-Model Integration at Half the Cost, Rivals Fable 5]** According to monitoring by Beating, OpenRouter has introduced the Fusion composite model interface, which allows user prompts to be distributed in parallel to multiple large language models, with the final answer synthesized through a referee model and a composite model. In the DRACO deep research benchmark test released by Perplexity AI, various model combinations based on Fusion demonstrated superiority over traditional single models. In specific configurations, combining Fable 5 and GPT-5.5 into a panel and synthesizing with Opus 4.8 achieved a top score of 69.0%, significantly surpassing the 65.3% performance of Fable 5 running as a single model. The test results indicate that due to differences in the underlying training and logic of models from different vendors, the diversity of multi-model combinations can provide stronger complementary perspectives in complex tasks. Even using the same model for self-combination showed a notable improvement: pairing Opus 4.8 with itself in a dual-panel self-synthesis setup increased the score from 58.8% (independent operation) to 65.5%. For cost-effective needs, a low-cost panel composed of Gemini 3 Flash, Kimi K2.6, and DeepSeek V4 Pro also achieved a score of 64.7% under synthesis, narrowing the gap with Fable 5 to just 0.6 percentage points while halving the invocation cost. The benchmark test covered 100 complex research tasks across 10 dimensions. To prevent models from inadvertently accessing online evaluation standards during retrieval, which could distort scores, OpenRouter has configured a filtered domain name list on the server side. Fusion is now available to API users, with the default identifier being `openrouter/fusion`. Users can also customize the model panels participating in synthesis via the web interface. [Original Link]