OpenRouter Fusion asks a panel of models, then a judge picks — 7x slower, 4x the cost
Fusion sends your prompt to several models at once, then a judge model synthesizes one answer. It scores well on a research benchmark. The thread's builders, who tried the same idea, explain why a judge mostly just rewards itself.
An engineer's quick eval put it plainly: about 7x slower and 4x the cost of calling one frontier model directly, which lands Fusion in a 'use it only when you really need it' bucket.
The deeper objection came from someone who shipped this exact thing months ago: asking one model to judge another mostly measures how closely the answer resembles what the judge would have said. It's temperature, dressed up as a panel.
What Fusion does
OpenRouter Fusion turns a single prompt into "a small multi-model deliberation": a panel of models analyzes it in parallel with web search enabled, then a judge model synthesizes their outputs into one response covering "consensus, contradictions, partial coverage, unique insights, and blind spots." There's a Quality preset and a cheaper Budget preset; pricing is the sum of the underlying calls. A reader supplied the missing context from the launch: routing to several models and combining their answers boosts performance on a deep-research benchmark, at the cost of time.
The cost-and-latency reality check
The first thing the practitioners did was measure it.
"I ran a quick eval to see what this looks like qualitatively vs just calling Opus 4.7 or GPT 5.5 directly. As expected, Fusion was 7x slower and 4x the cost. This isn't a knock against it… it places Fusion into a 'use it only when you need it' category." — a security-tooling founder who ran the eval
Where it pays off, another argued, is low-token, high-stakes work — analyzing a spec for gaps before coding, where "throwing as much LLM against it as possible is worthwhile regardless of cost."
"Asking one model to judge another" — a builder's autopsy
The sharpest reversal came from someone who'd already built and abandoned the idea.
"I built 'Fusion' a few months ago as an MCP… After extensive testing I discovered that when you ask one model to judge another's response you don't actually get a better answer. You are just asking it 'how closely does this resemble the answer you would have given me.'" — a developer who'd shipped the same concept
His verdict: extra rounds and the "obvious" fixes amount to "essentially just cranking up the temperature." A second builder reported the same against a frontier model: Fusion's output felt like "a bit diversified answer of the same class of pre-Fable SOTA models," missing the depth a single strong model reached. And one tester caught something odd in the logs — a billed call to a model that wasn't in their selection, surfacing only after the fact.
A panel is not a smarter member
Fusion is a clean implementation of an intuitive idea — many heads beat one — and on a research benchmark it shows. But the builders who've lived with model-juries keep landing on the same wall: a judge model can only reward what looks like its own reasoning, so ensembling raises diversity and cost more reliably than it raises correctness. The honest niche is the one the eval found: rare, expensive, high-value queries where you'd happily pay 4x to de-risk the answer — not your default chat box.
Get breakdowns like this every day
The top Hacker News stories, unpacked — by email, Telegram, or RSS.
Typst 0.15 ships multiple bibliographies and real HTML export — 'killed the invoice industry'
The LaTeX challenger's new release lands a feature researchers waited years for and pushes HTML/MathML export forward. The thread is full of people who already run businesses and write books on it — and one honest gap.