The interesting thing about OpenRouter’s Fusion isn’t the panel of models — it’s that the judge is the product.

Fusion runs your prompt across several expert models in parallel, with web search and fetch enabled, then a judge model synthesizes a structured analysis: consensus, contradictions, partial coverage, unique insights, blind spots. You’re billed for the sum of every panel completion plus the judge call. That pricing line is the honest part. Ensembling has never been free, and most teams who roll their own quietly eat the latency and token cost without ever measuring whether the extra calls changed the answer.

Having spent real time building automated micro-judges, the framing lands for me. The hard part was never “ask three models” — it’s the synthesis step. A judge that averages opinions is just a higher-variance single model. A judge that can name where the panel disagreed and which member was uniquely right is doing actual work, and that judge needs its own eval harness, not vibes. The moment you ship a deliberation layer, your correctness metric quietly moves from “is the answer good” to “did the judge weight the right member” — and those are not the same eval.

The blind-spots output is the one I’d instrument first in production. “Consensus” is seductive because it feels like confidence, but on a narrow task a confident committee can agree on the same wrong frame all day. The signal worth paying for is the disagreement, not the agreement.

The HN discussion already has the predictable “just cap it at two models” replies — which misses that the cost isn’t the panel, it’s whether your judge eval is good enough to justify any of them. So: has anyone actually measured judge accuracy separately from panel accuracy, or are we all trusting the synthesizer on faith?