Pinterest Slashes AI Costs 90% With Open-Source Model Strategy

Quick Facts

• Pinterest achieved 90% AI cost reduction by shifting from proprietary models to open-source alternatives
• Company reported Q1 2026 revenue of over $1 billion, up 18% year-over-year, with 631 million monthly users
• Pinterest built Canvas, an in-house AI image generation model that costs an order of magnitude less than third-party models

Pinterest cut its AI costs by 90% through a strategic shift away from expensive proprietary models to fine-tuned open-source alternatives. The visual discovery platform maintained similar performance levels while dramatically reducing operational expenses.

The company reported Q1 2026 revenue surpassing $1 billion, an 18% increase year-over-year, with 631 million global monthly active users. Pinterest repurchased roughly $2 billion of stock year-to-date, reducing shares outstanding by approximately 16%.

“In our quest to harness the power of AI, we were able to tap into available large-scale open-source models and achieve performance similar to proprietary models but at 90% less cost,” said CEO Bill Ready. “This addresses the return-on-AI-investment headache that many CEOs are facing.”

Pinterest adopted a hybrid approach to AI workloads. The company uses proprietary models for personalization, open-source models for cost-effective multimodal tasks, and closed-source models for high-performance use cases. Pinterest leverages OpenAI for some product features, Anthropic’s Claude for internal coding, and Alibaba’s Qwen for visual and content understanding.

The shift required significant technical innovation. Pinterest extended its proprietary generative retrieval system ‘Pennock’ globally in Q1, improving search fulfillment by 180 basis points while reducing cost per acquisition and cost per click by the same amount. The company updated its search ranking model to use up to 16,000 user actions over two years, a 30x increase in context that improved fulfillment by 70 basis points.

Pinterest built Canvas, an in-house AI image generation model trained exclusively on Pinterest data. The model operates at an order of magnitude lower cost than third-party alternatives while serving the company’s specific visual discovery needs.

“We’re a smaller company, so cost matters,” said Vicky Gkiza, VP of Product Management. “When discussing what drove this strategic shift for Pinterest, it came down to cost.”

The platform now serves hundreds of millions of inferences per second. Each user request completes thousands of model evaluations in under 100 milliseconds using hybrid CPU/GPU clusters.

Industry experts see Pinterest’s approach as part of a broader trend. “Companies are increasingly seeing value in a multimodal AI approach that balances performance and the cost of tokens,” said Lan Guan, chief AI and data officer at Accenture. “This token cost is going to slow you down if you don’t start managing them proactively.”

External analyses from a16z and SemiAnalysis confirm 10x to 30x inference cost reductions for some workloads by running optimized open models on reserved or owned GPUs versus retail API prices.

Pinterest reduced its workforce by less than 15% in January 2026 to reallocate resources toward AI-focused roles. The company expects modest cost headwinds from GPU and AI capacity investments, adding roughly 100 basis points of cost pressure.

Read more: Pinterest cut AI costs 90% by gutting a frontier model’s vision layer