Fusion Beats Frontier

OpenRouter's Fusion runs a research question through several model responses, then analyzes agreements, contradictions and missing coverage before producing a final answer. Its evaluation account suggests that some benefit can come from repeated attempts with the same model: two runs of Anthropic's Opus 4.8, synthesized by Opus 4.8, scored 65.5% against 58.8% for a solo run on 100 deep-research tasks from the DRACO benchmark.

The same article reports that mixed-model panels scored higher on that benchmark, for example 67.6% for Opus 4.8 with GPT-5.5, so the same-model result shows that synthesis alone may help rather than that model diversity is unnecessary. It measures the combined effect of an additional run and the synthesis step, and OpenRouter's FAQ says invoking Fusion is often two to three times longer than a standard call.