Cloudflare AI Search goes GA with image embeddings, OCR and pricing
- Source
- blog.cloudflare.com
- Author
- Cloudflare
- Date

Context
Cloudflare's AI Search, a managed service that indexes a developer's files and retrieves the relevant pieces for a search box or an AI answer, can now search images by what they look like and not only by a description. Until now, according to Cloudflare's general-availability announcement, AI Search detected objects in an image, wrote a caption and indexed only that text, so an image could be found only through words the caption happened to contain. It now also turns the image's pixels into an , a list of numbers positioned so that similar content lands close together, and keeps the caption alongside it. Cloudflare says this lets a search match details a caption leaves out, such as color, texture or layout. A user can also search with another image to find visually similar ones. Full image retrieval requires an embedding model that handles images, and Cloudflare names Qwen3-VL-Embedding as the one that supports it today. With a text-only model, an image query is first converted to a caption and searched as text.
The same release adds two other ways to get content in. Text files and PDFs can now be up to 10 MiB, up from 4 MiB. For scanned PDFs that contain only page images, an optional OCR step (optical character recognition, which reads printed text out of an image) extracts the text before indexing.
General availability also means AI Search becomes a paid service, with billing starting November 1, 2026. Cloudflare charges for three things: content ingested at $0.75 per million , storage at $2.00 per GB-month, and queries at $0.75 per thousand for semantic search or $0.10 per thousand for full-text search. Processing images, including OCR, is an add-on at $0.50 per million tokens. Parsing, , keyword indexing and are included, as is embedding with select Workers AI models; third-party models are billed separately. Every Workers plan gets a monthly free allotment of 5 million ingestion tokens, 10 GB of storage, and 1,000 queries of each type. Cloudflare changed one thing from the preview pricing: the free queries are now split into two pools of 1,000 rather than one shared pool of 2,000.
- embedding — A list of numbers representing a piece of text's meaning, so that similar meanings end up numerically close and can be searched.
- token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
- chunking — Splitting documents into passages small enough to embed and retrieve individually — the step that quietly determines whether retrieval works at all.
- reranking — A second pass that re-scores retrieved candidates by reading each one against the query, fixing the ordering that fast vector search got approximately right.
Checking sign-in…
Loading comments…






