Windmill is an AI-powered performance review platform that pulls context from HRIS systems, work tools like Linear, and manager observations to help companies make fair, data-driven people decisions. Their larger vision is a living context graph for every organization, continuously updated across every tool where work actually happens.

To run Windmill's own go-to-market operation, Ben Danzig, a software engineer at Windmill, built Pim: a fleet of 25+ distinct agents that run on a Mac Mini. These agents handle everything from deal lifecycles to advertising intelligence reports.

As Pim scaled, routing LLM traffic intelligently became essential. Not every task warrants the same model: some agents do sophisticated synthesis across the company's full data graph; others are classifying a deal stage change. Treating them the same would mean overpaying for every simple task and under-serving the complex ones.

To manage their LLM spend effectively, Windmill initially ran their LLM routing on OpenRouter. 

It worked, but when the Merge team introduced them to Gateway, the decision to switch over was easy for Ben. 

According to Ben:

“Merge Gateway offers better LLM routing, cost control, and observability functionality than OpenRouter. We only gained net-new capabilities by switching.”  

The process of transitioning over was equally compelling:

“Moving from OpenRouter to Merge Gateway was a single pull request. It just required an environment variable swap, a base URL change, and applying roughly 100 lines of code. This took minutes to implement.”

Windmill built to Merge’s Unified API for Gateway in a few hours. From there, they added dozens of models and implemented intelligent routing that optimizes for cost savings. 

Here’s just a snapshot of how this routing plays out in practice:

• Claude Sonnet 4.6 handles anything requiring multi-source synthesis or real-time judgment. For example, if a rep asks "what's the latest with this deal, I'm jumping on a call in ten minutes," the agent is pulling from CRM notes, Slack history, and call summaries and forming a coherent narrative on the spot
  

• Claude Haiku 4.5 covers high-frequency, low-complexity tasks, like deal stage classification, lead qualification, security and form processing. These are structured input/output jobs, so there’s no reason to pay for reasoning on a task that doesn't require it

• Kimi K2.5 manages weekly intelligence reports (e.g., advertising performance). Since this type of task processes large volumes of data on cron schedules rather than in real time, Kimi's large context window and competitive pricing make it the right fit

• GLM-5 powers the data visualization agent, like taking structured data, writing a dashboard, and deploying it to Vercel. Code generation against a well-defined schema doesn't require frontier reasoning, so GLM-5 can handle these tasks reliably at a cost that makes it practical 

On top of these routing rules, Windmill set a soft spending cap per month with alerts enabled. This ensures that the costs associated with Pim don't escalate unexpectedly and, if they do, Ben can review and address the root cause quickly.

According to Ben, Merge Gateway offers best-in-class governance and security features.

“Merge Gateway offers a centralized layer for policies, security guardrails, and searchable, request-level logs. This makes it easy to debug issues, stay audit-ready, and enforce consistent policies and controls across every model and team.”

In addition, the ability to consolidate several products to a single vendor has been super helpful.

“We already use Merge Unified to power our HRIS integrations. By using Merge Gateway, we can avoid juggling multiple vendor relationships and work with a team that already knows our product and our team," explains Ben.

Lastly, Ben can already cite tangible cost savings from adopting Merge Gateway.

“Thanks to Merge Gateway’s LLM routing functionality and spend cap rules, we’re able to save more than $10,000 per month on LLM spend without sacrificing performance.”

So, what’s next in Windmill’s plans to use Merge Gateway?

“We’re using our recent funding to build out our sales organization. This will expand Pim's scope and level of adoption, which will make Gateway even more important in the months ahead,” says Ben.

We can't wait to help them scale Pim, and keep the costs in check as they do.