Introducing Gumloop's Model Router

We decided to build Model Router after seeing both our team and our customers struggle to choose the right model. We noticed the same pattern: most of the time, people default to the smartest model available, regardless of the task.

On the surface, that feels like the right choice. After all, who doesn’t want the best output possible?

But not every task needs the latest frontier model like Claude Opus 5 or GPT-6 Astra. A routine daily brief doesn’t need the same reasoning as a complicated analysis.

And with new models coming out all the time, even our own team has trouble knowing which one to use.

At Gumloop, we’ve always believed the people closest to the work should be the ones creating agents to help automate it. They shouldn’t also have to become experts in model selection.

From $67K to $7K a month

We tested Model Router on one of our most important internal agents: our support agent. It was costing us over $67K every month.

After switching to Auto, it dropped to $7K.

We couldn’t believe it.

The agent was still handling the same support work, and we saw no difference in output quality. But our monthly cost was roughly 89% lower.

How it works

For each message, Model Router assesses the task’s complexity and how much reasoning it needs. It then chooses a model and reasoning effort to match, routing between open-weight and frontier models.

A simple recurring task might run on an efficient open-weight model like DeepSeek V4 Flash or GLM-5.3-Flash. More complex work might call for a frontier model like Grok 4.6, Claude Opus 5, or GPT-6 Astra, with the reasoning effort adjusted to match the task.

And if the work turns out to be harder than expected, the router can bring in a stronger model mid-run. So you’re not stuck with its first choice.

The result is better outputs, faster responses, and lower costs, without paying for more compute than the task actually needs.

Model Router is available for all users. Select Auto in the model picker and set it as your default.

Read the docs to learn more.