Cloudflare made AI Search generally available on 1 October 2026, adding native image embeddings, optical character recognition for scanned PDFs and support for larger files, according to its announcement. The company says the service combines its existing tools into a managed pipeline that indexes material and retrieves results for search or generated answers.
Key points
- Qwen3-VL-Embedding lets AI Search compare image pixels directly while retaining captions for text searches.
- OCR can read scanned PDF pages before indexing; the file limit rises from 4 MiB to 10 MiB.
- Cloudflare will begin billing on 1 November 2026, with monthly free allotments for ingestion, storage and queries.
Qwen3-VL-Embedding searches image pixels
Cloudflare says its earlier approach to image search detected objects, wrote a caption and turned that caption into an embedding, leaving the image searchable only through the details the description happened to include. AI Search now keeps the caption and can also embed the pixels directly using Qwen3-VL-Embedding. A query image processed by that model enters the same vector space as indexed images and text, allowing the service to look for visual similarities that the caption might have missed.
An embedding is a numerical representation that places related material near other related material for retrieval. Cloudflare says it uses Matryoshka Representation Learning so that smaller embeddings retain useful information, keeping storage manageable and search fast even with the richer representation of an image. The distinction matters in the company’s bird photograph example: a caption identifies the subject and surroundings, while the visual representation can preserve such details as the shape of a white eyebrow stripe and the texture of nearby fruit.
Finding a photograph with a particular eyebrow stripe could therefore rely on the markings in its pixels, even if its caption mentioned only the bird and the fruit around it. Matching it against another photograph could likewise depend on visual details rather than on whether the two descriptions used the same words.
Cloudflare says an image query still works when an AI Search instance uses a text-only embedding model, but follows a different route: ToMarkdown converts the image into text, and the service searches with the resulting caption. That provides a way to submit an image without the direct pixel comparison offered by a model with native image support. The company also says searches can combine an image with a written description, or use either on its own.
Workers AI and Vectorize run the retrieval pipeline
AI Search brings Workers AI, Vectorize, R2 and Browser Run together in a managed indexing and retrieval pipeline, Cloudflare says. For a query, the service can first rewrite the request and then embed it. Vector search, which compares embeddings, runs alongside keyword search, which looks for matching terms. The service combines those results, can rerank them, and either returns the leading chunks of material or passes them to a generation model to write an answer.
Cloudflare says it uses AI Search for search on its own blog and developer documentation. Cloudflare says it launched AI Search more than a year before the general-availability release.
OCR opens scanned PDFs up to 10 MiB
AI Search now accepts PDFs and text files up to 10 MiB, compared with the previous 4 MiB limit, Cloudflare says. The supported text formats include Markdown, HTML, CSV and JSON. For a PDF made up of scanned pages, turning on OCR lets the service read the text on each page before dividing that text into chunks and embedding it for retrieval. Without that reading step, the words in a scanned page are part of an image rather than extractable PDF text.
Cloudflare says OCR is available to every account and is charged as image-processing ingestion tokens. That charge is relevant to the broader visual-search release: the company says ingestion uses the same base token rate regardless of which Workers AI embedding model is chosen, but processing images adds a separate fee. A change from a text-only model to a multimodal one therefore has a different pricing effect from adding image processing to the material being indexed.
AI Search billing begins on 1 November
Cloudflare will start billing for AI Search on 1 November 2026, using prices it announced in preview during its August 2026 Agents Week. The company sets base ingestion at $0.75 per 1M tokens and image processing at an additional $0.50 per 1M tokens. Stored data costs $2.00 per GB-month. Semantic queries, including hybrid and vector searches, cost $0.75 per 1k queries, while full-text queries cost $0.10 per 1k.
The free monthly allotment on all Workers plans covers 5M ingestion tokens, 10 GB of storage, 1,000 semantic queries and 1,000 full-text queries, Cloudflare says. Those query allotments replace the shared pool of 2,000 queries in its preview pricing. The ingestion allotment is a single pool covering supported file types, including text and images, rather than a separate allowance for each.
Cloudflare says its charges cover the content indexed, the data stored and the queries run, with parsing, chunking, keyword indexing and reranking included. Embedding and reranking are free with select Workers AI models; third-party models are billed separately. The company says the current keyword-search implementation has limits with big data stores and is refactoring that engine to scale better.
