OpenRouter launched Fusion this week. The headline is that it matches Claude Fable 5 on a research benchmark for about half the cost, and they're calling it the smartest compound model on the market.
I read the whole thread, and the thing that stuck with me wasn't the headline. It was a number they dropped halfway down. They say roughly three quarters of Fusion's improvement comes from the synthesis step, and only about a quarter from using a mix of different models.
That's a strange thing for OpenRouter to admit. Their business is selling access to lots of models. And here they are telling you the models are the minor ingredient.
How it actually works
You send a prompt. Fusion sends it out to several models at once. A “judge” model reads all the answers and works out where they agree, where they contradict each other, and what each one caught that the others missed. Then a “synthesizer” model writes the final answer off the back of that.
So the part that moves the needle, by their own measurement, is the judging and the writing-up at the end. Not the panel itself. I find that genuinely interesting, and a little awkward for them.
The cheap-models result
This is the part I keep coming back to. OpenRouter says a panel of budget models (Gemini 3 Flash, Kimi K2.6, and DeepSeek V4 Pro) came within about 1% of Fable 5 once they were fused together. That same panel apparently beat solo GPT-5.5 and solo Opus 4.8 outright.
If that holds up, it suggests raw model quality matters less than I'd assumed. You might not need the best model in the room. You need a few decent ones and a good way to combine what they say.
I'm hedging on “if it holds up” for a reason. This is one benchmark, run by the company selling the thing.
Where I get skeptical
Two things bug me.
The price claim. “Half the cost of Fable” reads clean, but you're not running one model anymore. You're running a whole panel of them, plus a judge, plus a synthesizer, on every single query. Cheaper than Fable for comparable quality on this test, maybe. It's still a lot more compute and a lot more waiting than a single model call, and the “just call one slug” framing skips past that.
The benchmark. They tested on DRACO, a deep-research benchmark from Perplexity. Deep research happens to be close to the best possible case for this approach, because the actual work is reconciling a pile of sources. I'd want to see how Fusion does on quick factual questions, on coding, on anything creative, before I bought “smartest compound model on the market” as a general statement. On a one-line question, fanning out to five models and running a judge over them is probably just slower and pricier for no real gain.
What I take from it
For the last couple of years the conversation has been about the model. Which one scored highest, which lab shipped what. Fusion's own data points somewhere else, up a layer, into how you route a question and combine the answers you get back.
I don't think this means models stop mattering. A panel of bad models won't save you. But OpenRouter set out to sell a smarter model and ended up making a fairly good case that the answer matters less than what you do with it afterward. I doubt that was the plan.
Opinion — Jlabs Research. Sources: OpenRouter Fusion announcement and company benchmarks; DRACO deep-research benchmark (Perplexity). Figures are relative illustrations; OpenRouter did not publish per-model scores. Not investment advice.