Merge Launches Fusion: Multi-Model Ensemble Claims Frontier Performance at 1/4 Cost

Merge launched Fusion, turning traditional machine learning ensemble methods into a production-grade API primitive. With a single API call, the system sends the same prompt to multiple models, and a judge model integrates the outputs to generate the final answer. The official claims state that this solution beats Fable 5 at 1/4 the cost and achieves the highest score on the DRACO benchmark, marking a shift in LLM usage from "choosing the smartest single model" to "multi-model collaboration with adjudication."

Confirmed

Fusion is now live on Merge Gateway, supports mainstream models, and offers a $10 free credit across all plans. Official supplementary benchmark results show that all Fusion configurations outperform standalone frontier models. The tests covered 25 research questions, demonstrating that even the cheapest configuration using only open-source models achieves frontier-level quality in accuracy, completeness, and objectivity. Incorporating stronger models further expands this advantage.

Why it matters

@sull notes that the core significance of Fusion lies in converting ensemble methods into a directly callable production API primitive. If the claimed cost and performance data hold up to independent verification, this paradigm of parallel multi-model execution followed by judge-based adjudication could profoundly change how developers use and orchestrate LLMs.

2026-07-23 ~ 2026-07-24 · 12 related posts

Primary sources

5 near-duplicate retellings: thisguyknowsai · thisguyknowsai · thisdudelikesAI · bigaiguy · thisdudelikesAI