Pinterest cuts AI costs 90% by replacing frontier vision layers with proprietary embeddings

Pinterest says it has cut the cost of some AI-powered experiences by 90% by stripping back a frontier model’s vision layer and replacing it with its own proprietary embeddings, a move that underscores how major apps are trying to make generative AI cheaper to run at scale. The company’s CTO, Matt Madrigal, described the approach as part of Pinterest’s broader effort to customize open-source models “foundationally in-house,” according to VentureBeat.

The change matters because Pinterest operates at enormous scale, with about 620 million monthly users, making expensive model calls unsustainable for routine image recommendations, as VentureBeat reported. Instead of relying on a frontier multimodal model for every visual query, Pinterest rebuilt part of the system around its own data and embeddings, which the company says not only reduced costs sharply but also improved accuracy by 30%.

The strategy fits a wider shift in how Pinterest is building AI products. Business Insider reported that the company has adopted a “model-agnostic” approach that mixes proprietary systems, closed models from companies such as OpenAI and Anthropic, and open-source models from Alibaba. According to that report, Pinterest has been using open-source AI for visual and content understanding, data labeling, and assistant tasks, while reserving more expensive models for areas where they still perform best.

That hybrid architecture is part of why Pinterest has been able to launch new AI features without letting infrastructure costs spiral. Business Insider said the company has used this blended model stack to support features rolled out in 2025, while also telling investors that AI-driven experiences can cost nearly 90% less than when Pinterest relied only on proprietary models. TechCrunch, in a separate report, framed the same cost discipline as increasingly important across the AI industry, describing budget cutting as a major selling point for companies selling AI tools.

For Pinterest, the commercial logic is straightforward: visual search and recommendation systems depend on large amounts of image understanding, but not every task requires the most expensive model available. By tuning open-source systems around its own proprietary data, the company is trying to preserve performance while reducing the recurring costs of inference, the compute used every time a model responds.

The broader significance goes beyond Pinterest. The company’s approach reflects a growing industry lesson that AI success may depend less on using the biggest model and more on re-engineering models around a company’s own data and use cases. As reported by VentureBeat and Business Insider, Pinterest’s executives are betting that this kind of hybrid, customized stack will let the company keep expanding AI features while keeping the bill under control.