Ceramic AI wants to slash web search costs from $5–15 to 5 cents per thousand queries

Summary

Anna Patterson, founder of Ceramic AI and former Google VP of Engineering, is targeting a structural cost problem in AI applications. Search costs 10 to 30 times more than inference, and that gap is blocking the adoption of agentic workflows.

Inference now costs roughly $0.50 per thousand queries. Search runs $5 to $15 per thousand queries. Patterson compares inference to a meal and search to a condiment—paying $500 for condiments when the meal costs $5 makes no sense, especially when agents need to search constantly to avoid hallucinating.

Ceramic plans to cut search costs to 5 cents per thousand queries, a 100x reduction from the high end. The company operates a 40 billion-page web index and uses supervised generation. As a model generates output, it continuously checks itself against live search results and asks what else should be included to stay comprehensive. This approach reduces hallucinations while maintaining low latency. One customer running agentic workflows achieves 50-millisecond response times across 1,100 searches.

Patterson describes a two-track go-to-market approach. Land large players seeking cost advantages, and build self-serve APIs for startups. For privacy-sensitive enterprises in pharma and banking that need internal search indexes, offer custom deployments.

Ceramic respects robots.txt restrictions and doesn't scrape blocked sources. Patterson says the company is in active talks to strike revenue-sharing deals with publishers who want more control over their content. This points toward a longer-term model where proprietary content comes with premium APIs at higher cost to customers, with revenue flowing back to publishers.

Patterson also notes that spare compute in modern mixture-of-experts models could enable new product categories. Push-based alerts powered by cheap search could run continuously, streaming fresh signals and filtering out information the user already knows.