An internal incubator for AI products and tools within the Meta Platforms, Inc. ecosystem is developing a routing service designed to compete with OpenRouter.
This service aims to distribute certain AI tasks to lower-cost models, thereby reducing overall computational expenses.
The incubator, known as AAI Labs, falls under Meta's Applied AI Engineering division.
Established earlier this year, this division allows employees to propose and develop their own AI product and tool concepts for internal use first, with potential for future public release.
Internal documents indicate that once a project proposal is approved, a small dedicated team is formed to complete development, with a view to potentially launching it to the market.
A July internal memo reveals that AAI Labs currently has approximately 200 approved projects, spanning three main areas: consumer applications, developer tools, and internal enterprise infrastructure.
The AI model routing tool, named Switchboard, is one such project.
This system evaluates the complexity of requests from users or AI agents and automatically assigns them to appropriate models.
Simple interaction tasks are routed to lighter, more cost-effective models, with a logic similar to OpenRouter's Auto Router product.
Sources familiar with the matter indicate that, like many projects within AAI Labs, Switchboard is currently in an early development phase and may never be publicly released.
However, the project team's proposal document outlines its dual value: internally, it could help Meta cut costs, and externally, it could be a commercially viable product for enterprises deploying code agents at scale.
The portfolio of projects incubated by AAI Labs reflects Meta's core objective: leveraging its massive AI investments to explore new tools, commercial ventures, and revenue streams beyond its primary advertising business.
Meta estimates its total expenditure on AI computing infrastructure and hardware this year could reach $145 billion, more than double the projected 2025 spend.
The company is also continuously adjusting its engineering organization to bolster AI R&D capabilities.
By empowering internal employees to incubate new projects, Meta can rapidly prototype AI tools that can both optimize internal operations and be refined into standalone commercial products.
Furthermore, Meta has been actively managing the significant costs associated with employee use of AI development tools.
Following a company-wide push for AI tool adoption earlier this year, Meta quickly implemented token usage limits and built an internal control platform to track AI costs and enforce budget management.
Industry Significance of AI Model Routing
OpenRouter gained popularity among developers by offering a single point of access to multiple large language models, helping control costs and improve efficiency.
Recent reports suggest a major tech company is in talks to acquire OpenRouter, a deal that could further increase its valuation, which stood at $1.3 billion as of April.
The model routing technology sector gained widespread industry attention in 2025 after OpenAI's GPT-5 introduced built-in model switching for simpler queries.
Companies like Databricks and Palantir have also developed their own routing tools for cost control and operational efficiency.
Meta's internal documents highlight the specific problem Switchboard aims to address: "Even for the simplest coding tasks, we are paying as if we're using the most advanced, high-end models."
The project proposal notes that the vast majority of code agent tasks can be handled by mid-sized or smaller models, with only a few complex tasks requiring expensive, cutting-edge models.
The current practice of routing all tasks to a single model means overpaying for simple tasks while complex tasks may not receive sufficient computational resources.
The documents state that high inference costs are the primary obstacle to the widespread internal deployment of AI agents at Meta, as cost ceilings limit the scale of deployment.
Meta declined to comment on this report.
AAI Labs is also developing another product: an AI-powered voice-guided tour application for drivers, compatible with Apple CarPlay and Android Auto.
According to planning documents, this product uses AI to provide real-time narration about landmarks and allows drivers to ask for related information.
Positioned as an extension of Instagram's map feature, future plans include integrating location-based short videos (Reels) and travel recommendations, with potential connectivity to Meta's Ray-Ban smart glasses.
The app will be tested internally with employees before a potential public release.
The establishment of AAI Labs aligns with CEO Mark Zuckerberg's broader strategic vision that AI enables small teams to develop products more rapidly.
During an April earnings call, Zuckerberg told analysts that AI agents empower small teams to achieve efficient development and predicted AI would spur a wave of innovation, suggesting Meta could launch up to 50 new proprietary applications in the future.
Comments