Meta’s AI Incubator Is Developing an OpenRouter Rival to Cut Coding Costs
The Takeaway
- Meta’s new AAI Labs incubator is developing an OpenRouter rival called Switchboard
- Switchboard is among 200 early-stage projects at AAI Labs
- Router would send some AI tasks to cheaper models to reduce inference costs
Meta Platforms’ internal incubator for AI-powered products and tools is developing a version of the OpenRouter service that would help cut costs by sending some AI tasks to lower-cost models.
The incubator, called AAI Labs, is part of Meta’s Applied AI Engineering team, which Meta set up in March and allows employees to pitch AI-powered products and services for internal use—and possibly later release them to the public. Once a proposal is approved, a small team is assembled to build and potentially launch it, according to internal documents reviewed by The Information.
A July memo says AAI Labs has about 200 approved projects spanning consumer products, developer tools and internal infrastructure.
The AI model router, dubbed Switchboard, is among those projects. It would determine which models should handle each request from a human user or AI agent by scoring tasks by difficulty. It would send simpler requests to smaller, cheaper models, similar to the way that OpenRouter’s Auto Router product works, according to one of the documents.
As with many projects inside AAI Labs, Switchboard is an early-stage effort and may never be launched as a product, according to a person familiar with the matter. But the Meta team working on the project said in the document outlining the project proposal that the company could use Switchboard internally to lower costs, and release it publicly to organizations running AI coding agents at scale, the document shows.
Projects developed under AAI Labs reflect Meta’s appetite for turning its enormous AI investment into potential new tools, businesses and revenue streams beyond advertising. The company has projected its spending on AI infrastructure and other equipment and facilities could hit $145 billion this year, more than double the amount in 2025, and it has been reorganizing its engineering teams to strengthen AI development.
The AAI Labs projects also show how Meta is turning to employees for this effort, to quickly prototype AI-powered products that can either improve Meta’s internal operations or become standalone products released to the public.
Meta also has been looking for ways to rein in the billions of dollars it is spending on AI tools for its coders and other employees. As The Information previously reported, the company told employees in June that it would begin imposing limits on AI token usage just weeks after encouraging broader adoption of AI tools across the company, while also building an internal platform to track AI spending and enforce token budgets.
The Value of AI Routers
OpenRouter has gained popularity among developers looking to access various AI models while trying to reduce costs. The Information last week reported that the company has held discussions about a potential acquisition by a bigger technology company—a deal that could boost OpenRouter’s valuation by billions of dollars. In April, it was valued at $1.3 billion valuation.
Model routing gained broader attention last year when OpenAI released GPT-5, which included a router that automatically switches to a cheaper model when a user’s prompt is relatively less complex. Other companies, including Databricks and Palantir, have developed their own router tools to help manage costs and improve efficiency.
One of the internal Meta documents describes the problem that the company’s engineers hope to address through Switchboard: “We pay top-model prices for every coding request, including the easy ones.” Most coding-agent tasks can be handled by smaller models, while only a minority require the capabilities of more expensive frontier models, according to the project proposal rationale outlined in the document. “Today everything goes to one model, so we overpay on easy work or underperform on hard work,” the document says.
The document states that inference costs are the primary barrier to deploying agents more broadly across the company. “Cost is what limits how widely we can run agents,” the document says.
Meta declined to comment.
Among the other projects under development at AAI Labs is an AI-powered tour guide app for drivers that would run through Apple CarPlay and Android Auto. It envisions using AI to narrate nearby landmarks and to allow drivers to ask questions about those points of interest, according to a separate internal document outlining the plans. The document positions the product as an extension of Instagram’s Map experience that could eventually incorporate location-based Reels and travel recommendations and perhaps later integrate with Meta’s Ray-Ban smart glasses. Meta could release the app publicly after testing it among employees first, according to the document.
The formation of AAI Labs aligns with CEO Mark Zuckerberg’s broader vision that AI will enable smaller teams to build products more quickly. In April, he told analysts that AI agents mean “small groups of people and teams can make very rapid progress” and predicted the technology would drive “a lot of innovation.” Zuckerberg also has said that Meta could build as many as 50 new apps.