Cloudflare makes AI Search generally available, with multimodal retrieval and billing starting November 1

A team picking a retrieval layer for an agent has to decide what it is willing to run itself, and the cost of the alternative just dropped onto one page.

Cloudflare announced on 1 October that AI Search, its managed retrieval product, is generally available, and that it will start charging for the service on 1 November 2026. The company says the GA launch adds native multimodal retrieval using the Qwen3-VL-Embedding model, OCR on PDFs so scanned documents can be searched, and a larger maximum file size for ingestion.

AI Search is the hosted layer sitting between a team's documents and an agent that needs to look things up: it ingests files, chunks and embeds them, stores the vectors, and serves either semantic queries or full-text queries from Workers, external code or a Model Context Protocol client. Cloudflare says it launched the product more than a year ago, with preview pricing announced during Agents Week in August 2026. The GA launch sets fixed prices and a free tier, which is what a team evaluating it against a self-hosted stack of Postgres with pgvector, Pinecone, Weaviate or a Vectorize-plus-R2 pipeline needs in order to run the numbers.

The published prices, in Cloudflare's own words, are $0.75 per million tokens for base ingestion, an additional $0.50 per million tokens when image processing is turned on, $2.00 per GB-month for stored data, $0.75 per thousand semantic queries and $0.10 per thousand full-text queries. The free monthly allotment covers 5 million ingestion tokens, 5 million more for image processing, 10 GB of storage, 1,000 semantic queries and 1,000 full-text queries. Cloudflare says it will not start billing until 1 November, so there is a month to use the service without charges.

The multimodal change is what will matter to teams whose documents are not plain prose. The preview could retrieve images only through captions, so a search for a chart was really a search for the caption describing it, and anything the caption did not say was invisible. With native multimodal retrieval, Cloudflare says the Qwen3-VL-Embedding model embeds the image itself alongside its caption, so a query can match on visual similarity, texture, composition or layout as well as on the text that happens to sit next to the image. The company says OCR on PDFs reads the text from each page before chunking and embedding, so scanned contracts, financial statements and research papers can be searched the same way as a native text document.

For a team that is already on Cloudflare and wants retrieval without running a vector store, the main thing to work out between now and 1 November is which part of the bill grows fastest with real usage. Ingestion is a one-time cost per document version, so it falls as content stabilises. Storage is linear in corpus size and queries scale with traffic, so a product doing semantic search inside a chat interface will see semantic-query cost dominate, while a documentation search box will see full-text-query cost dominate. The Cloudflare announcement does not publish a comparison against any other retrieval product. The useful exercise before 1 November is to point AI Search at a real corpus, measure its hit rate on a frozen set of production questions, and compare that to whatever is answering those questions today.

Source

• Cloudflare, AI Search is now generally available, 1 October 2026