# Merge launches Fusion to combine multiple AI models into one answer > Source: > Published: 2026-07-23 16:20:06+00:00 [Merge](http://merge.dev/gateway) co-founder and CEO [Shensi Ding (@shensi)](https://x.com/shensi) launched Fusion on July 23rd, giving developers one API call that sends a prompt to several AI models and uses a separate judge model to assemble their responses into a final answer. Ding announced the product [in a thread on X](https://x.com/shensi/status/2080317858551124260), framing Fusion as a way to beat the quality of a single frontier model without paying frontier-model prices. The release pushes Merge deeper into AI infrastructure, six years after Ding and co-founder [Gil Feig (@GilFeig)](https://x.com/GilFeig) started Merge to handle a different type of software fragmentation. Ding and Feig met while studying computer science at Columbia University. They later encountered the same integration problem from opposite sides of two San Francisco startups: Ding at cybersecurity business Expanse and Feig as head of engineering at recruiting software maker Jumpstart. They [started Merge in June 2020](https://www.merge.dev/blog/why-we-started-merge) to let software developers connect to business applications through a unified API. Fusion applies that original aggregation thesis to the increasingly fragmented model market. Developers specify at least two analysis models, which receive the full prompt in parallel, and can select a separate synthesis model to judge the candidates. The result arrives through Merge Gateway's standard responses endpoint, accompanied by metadata identifying which models participated and how many completed the request. The [Fusion documentation](https://docs.merge.dev/merge-gateway/features/fusion) makes the tradeoff explicit. Each request incurs a call to every panel model and another call to the judge. A three-model panel therefore generates roughly four billable model calls. Fusion is also non-streaming, and Merge recommends it for work where answer quality outweighs response time, such as research and analysis. That cost structure is central to Merge's pitch. Fusion does not reduce the amount of inference performed. It tries to substitute several lower-priced calls for one expensive frontier-model call, then relies on the judge to preserve the strongest parts of each response. ### Merge's benchmark claim In a [technical post accompanying the release](https://www.merge.dev/blog/merge-fusion), Merge said it evaluated Fusion on 25 tasks selected across the 10 domains in DRACO, a deep-research benchmark created by Perplexity researchers. The [full DRACO benchmark](https://research.perplexity.ai/articles/evaluating-deep-research-performance-in-the-wild-with-the-draco-benchmark) contains 100 tasks derived from anonymized user research requests and grades responses for accuracy, completeness, presentation and citation quality. Merge said every Fusion configuration in its test outscored every individual model tested. According to Merge, a panel composed entirely of models it describes as open source scored statistically level with Anthropic's Claude Fable 5 while costing about one-quarter as much. A more expensive panel beat Fable 5 by 8.5 points, Merge said. Fusion completed all 25 selected tasks in two runs, while the solo frontier model refused the same task twice. Those results remain a Merge-run product benchmark. The published announcement covers one-quarter of DRACO's full task set and does not include raw task-level responses or an independent replication. The numbers show how Merge configured Fusion for its own test; developers will still need evaluations based on their prompts, latency limits and failure costs before treating a panel as a replacement for a frontier model. Fusion also gives customers control over which models sit in the panel and which model acts as judge. That flexibility prevents Merge from offering one universal quality or cost result. Changing a panel member, allowing web search, increasing output limits or choosing a premium judge changes the bill and can change the answer. ### A gateway becomes the product Fusion is part of a broader repositioning at Merge. Ding wrote that Merge made a hard internal shift during 2025 toward AI use cases and products. Merge now sells three layers: Unified for synchronized business data, Agent Handler for tool access, and Gateway for model routing, fallback, security and cost controls. That expansion gives Merge a larger role in each AI request. Fusion brings model selection and answer synthesis inside Gateway rather than leaving orchestration code with the customer. It also increases routed model volume, which matters because [Merge Gateway's Pro plan](https://www.merge.dev/pricing/gateway) charges underlying LLM costs plus a 5% fee. Merge offers $10 in monthly credits on that plan, while enterprise pricing is negotiated. Merge has raised about $74.5 million. Its most recent disclosed round was a [$55 million Series B](https://www.merge.dev/blog/announcing-merges-55-million-series-b) led by [Accel](https://www.accel.com/companies/merge), with participation from NEA and Addition, in October 2022. Fusion is available across the major models carried by Merge Gateway. Developers can start with $10 in credits, according to Merge. The product's immediate constraint is latency: every panel member must answer before the judge can synthesize the result. For research agents and other workloads where one stronger response can justify waiting, Merge is betting that orchestration can compete with buying the most expensive model directly.