/TECHStory update pending

OpenRouter launches Fusion API, a compound multi-model system claiming deep research capabilities at half the cost of monolithic models

It outperforms solo frontier models on the DRACO benchmark.

Story Brief

It outperforms solo frontier models on the DRACO benchmark.

Commentary on X

Highest ranked

@yacineMTB But it's compared to frontier models solo. So a bit apples to oranges. These models are now meant to do compounding as well. Taking that away, still scoring below, and calling it smarter, is a kind of BS that will not go unnoticed and plays against their credibility.

Laura@laura_bh4c03lTECH

View all

RipVanWinkle@twplsourceTECH

@iamtrask awesome!!! 👏 👏👏

⿻ Andrew Trask@iamtrask

Sentiment

Positive68.1%31.9%Negative

Positive Read

Many users praised OpenRouter's Fusion API for routing prompts across models to deliver frontier intelligence at half the cost, while others called the claims misleading and accused it of overcharging or scamming in practice.

Based on 339 visible X reactions from 1262 accounts.

Digg Deeper

so, the tldr is that the same request hits multiple models then a judge sorts through the mess, the judge uses a synthesizer, who wrote the synthesizer and what model is it using?
OpenRouter's team (CEO Alex Atallah and engineers) built the Fusion system, including the judge/synthesizer orchestration and prompts.[[1]](https://x.com/i/status/2065856853989270011) It fans out the prompt to a panel of models (with tools), has a judge extract structured analysis (consensus, contradictions, gaps, insights), then a synthesizer produces the final grounded answer; users can override the judge/synthesizer model via the API, but the default isn't named in the announcement or docs snippets.
1
this will suck at coding, correct? no way multi-models + judge + synthesizer is fast/produces coherent code
No coding-specific claims or benchmarks exist for Fusion API. It is positioned for deep research tasks, where multi-model panels deliver Fable-level results at half the cost. The judge-plus-synthesizer routing across diverse models inherently increases latency and risks coherence issues versus a single strong model for code.