Pinterest cuts AI costs 90% by mixing open-source and proprietary models

Pinterest cuts AI costs 90% by mixing open-source and proprietary models

Pinterest is shrinking its AI budget while expanding AI features across its platform. The company achieved a 90% cost reduction by combining proprietary models it built in-house with free open-source options and expensive closed-source models from OpenAI and Anthropic, according to Vicky Gkiza, the company's vice president of product management.

The strategy, which began in 2023, treats different AI models as tools for different jobs rather than betting on a single approach. Closed-source models handle tasks requiring speed and minimal maintenance. Open-source models, which are free to download and modify, handle cost-sensitive work.

How the model mix works

Pinterest uses OpenAI's models for product features and Anthropic's Claude for internal coding tasks. For visual understanding, content labeling, and assistant functions, the company relies on Alibaba's Qwen, an open-source large language model.

This blend powered two major feature launches in 2025. Auto-collages lets advertisers convert product catalogs into pins for shopping feeds. A voice-enabled search feature lets users ask questions aloud instead of typing, which early testing showed increased shopping-specific queries.

The company tested auto-collages starting in early 2024, then piloted it with retailers like Macy's by June 2025. Voice search entered beta testing in October 2025.

The talent investment

Executing this approach required hiring engineers who could customize and maintain open-source models. Pinterest brought in Matthias Zenger, a former Google engineer, as vice president of engineering in April 2025. Three months later, the company hired software engineer Mirjam Wattenhofer and announced an Engineering Excellence Hub in Zurich.

Both work on improving user experiences with AI and machine learning. In February, CEO Bill Ready said Pinterest would hire additional research and development staff to support AI efforts.

Gkiza said the company is "investing more in hiring the right talent - evolving the team, whether it is engineering or product management - to be much more familiar with AI."

Infrastructure and next steps

Open-source models require computing power that proprietary models abstract away. Pinterest plans to invest more in cloud infrastructure, particularly graphics processing units, to run these cost-saving systems.

The company will continue testing different models, prioritizing its own for personalization, open-source options for cost and flexibility, and closed-source models when they deliver superior performance.

Lan Guan, chief AI and data officer at Accenture, said companies increasingly see value in this mixed approach. "This token cost is going to slow you down if you don't start managing them proactively," Guan told Business Insider. "Open-source will be a really good option."

For product teams, the lesson is clear: the most expensive option isn't always the right one. Generative AI and LLM strategies that match model type to use case can reduce costs without sacrificing performance. AI for Product Development increasingly means knowing when to build, when to buy, and when to use free tools.