OpenRouter's Fusion stitches AI models together for better answers at half the cost

OpenRouter just launched Fusion, a composite model interface that shoots your prompt to multiple large language models in parallel, then uses a judge model and a synthesizer to spit out a final answer. In Perplexity AI's DRACO deep research benchmark, Fusion-powered combos crushed traditional single models. The winning setup — a panel of Fable 5 and GPT-5.5 synthesized by Opus 4.8 — scored 69.0%, comfortably beating Fable 5's lone 65.3%. The diversity of different vendors' training and logic apparently brings stronger perspective on complex tasks.

Even pairing a model with itself helps: Opus 4.8 running solo scored 58.8%, but when it formed a dual panel with itself and synthesized, it jumped to 65.5%. For the cost-conscious, a low-budget panel of Gemini 3 Flash, Kimi K2.6, and DeepSeek V4 Pro managed 64.7% — just 0.6 points behind Fable 5 while costing half as much.

The benchmark includes 100 complex research tasks across 10 dimensions. To prevent models from accidentally accessing online evaluation standards and skewing results, OpenRouter filters retrieval domain lists server-side. Fusion is now available to API users under the default identifier openrouter/fusion, and you can customize the model panel on the web.