Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarCloudflare AI Blog→ original

Cloudflare Takes AI Search Out of Beta: Image Search and Scanned PDFs, From $0.75 per Million Tokens

Cloudflare has opened AI Search to everyone, a ready-made service that splits files, indexes them and searches them, so you don't have to assemble a RAG pipeline from five separate parts. It now finds images by their pixels rather than only by their caption, and it reads scanned PDFs with OCR, the text recognition that pulls words out of page images. Starting November 1, 2026, uploads cost $0.75 per million tokens (a token is a chunk of a word). Estimate your document volume in advance so the bill doesn't catch you off guard.

AI-processed from Cloudflare AI Blog; edited by Hamidun News
Cloudflare Takes AI Search Out of Beta: Image Search and Scanned PDFs, From $0.75 per Million Tokens
Source: Cloudflare AI Blog. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceCloudflare AI Blog ↗
Editor-in-Chief
Storage$2 per GB per month, with the first 10 GB free on Workers
Queries$0.75 per 1,000 semantic, $0.10 per 1,000 full-text
Key metric #3File size limit raised from 4 to 10 MiB; images are searched with Qwen3-VL-Embedding

If you have built a documentation search or a RAG system (a setup where a neural network answers from your own files), you know how much has to be stitched together: a file parser, a chunker, an embedding model, a vector database, keyword search, and a reranker. Cloudflare offers all of this as one managed service, AI Search, which has now officially left beta. What's new: images are searched by their pixels rather than by caption text; OCR (optical character recognition) reads scanned PDFs, where the pages are photographs instead of text; and the service becomes paid on November 1, 2026, with file uploads costing from $0.

75 per million tokens. A token is a piece of text about the size of a syllable or a short word, so a million tokens is roughly a thousand to fifteen hundred pages.

What's inside and what's new

AI Search combines Workers AI, Vectorize, R2 and Browser Run into a single pipeline: you upload files or connect a site, and the service parses them, indexes them and answers queries. Cloudflare uses it to search its own blog and developer documentation.

The main update concerns images. Previously, a picture was first described in text: objects were detected, a caption was generated, and only the caption was indexed. Search worked, but only on the details that made it into the caption. Say you have a clothing catalog. The caption "blue jacket" says nothing about the fabric texture, the cut of the collar or the exact shade. Now AI Search stores both signals: the caption and an embedding (a numerical "fingerprint") of the pixels themselves.

Native image search currently runs on the Qwen3-VL-Embedding model. A query can be text, another image, or both together, for example "bird with similar plumage." The service checks whether your embedding model can work with images. If it can, the query image lands in the same vector space as the indexed texts and images. If the model is text-only, a query can still be sent as an image: ToMarkdown converts it into a description, and the search runs over that description. Basic multimodality is available with any model, but only models with image support get the full visual signal.

To keep vectors from inflating storage, Cloudflare uses Matryoshka Representation Learning. With this approach, a shortened vector retains most of the useful information, so search stays fast and storage stays cheap.

The query path works like this: the query is optionally rewritten and turned into a vector. Vector and keyword search run in parallel, the results are merged and, if you choose, reranked. You then receive the best fragments or pass them to a model that writes the answer.

Two more practical changes. The limit for text files (Markdown, HTML, CSV, JSON) and PDFs has grown from 4 to 10 MiB. For scanned PDFs, OCR can be enabled: the service reads the text from each page before chunking and indexing.

How much it costs

Billing starts on November 1, 2026, and Cloudflare promises an email reminder beforehand. You pay for three things: content uploads, storage and queries. Parsing, chunking, embedding with Workers AI models, keyword indexing and reranking are included in the price. There are no instance-hour charges and no monthly minimums.

The rates are:

  • uploads: $0.75 per million tokens; image processing (including OCR) adds $0.50 per million;
  • storage: $2 per GB per month;
  • semantic queries (vector and hybrid): $0.75 per 1,000;
  • full-text queries: $0.10 per 1,000.

Every Workers plan includes these free each month: 5 million upload tokens, 10 GB of storage, and 1,000 semantic and 1,000 full-text queries. The last item is a recent change: the preview had a shared pool of 2,000 queries. The Workers AI models selected for embeddings and reranking are free; third-party models are billed separately.

For a rough estimate (my own calculation from the published prices): 20 million tokens of documentation and 50 thousand semantic queries per month would cost about $48. Of that, about $11 goes to uploads and the rest to queries, and a 5 GB index fits within the free storage allowance. Switching from a text embedding model to a multimodal one does not change upload costs unless you process images.

Limitations and plans

Video and audio are not supported yet; Cloudflare is only preparing a pipeline for them. The company is rewriting keyword search, because the current implementation has limits on large indexes. Separately, the team is working on simpler connection of AI Search to sites on Cloudflare, so AI agents can find and read content more easily.

The practical takeaway: if your search revolves around catalogs, screenshots, diagrams or scans, it is worth testing native image embeddings, which capture what didn't fit into a caption. And before November, estimate your token, storage and query volumes so the first bill doesn't come as a surprise.

Story

Key Questions Answered

What's inside and what's new?

AI Search combines Workers AI, Vectorize, R2 and Browser Run into a single pipeline: you upload files or connect a site, and the service parses them, indexes them and answers queries. Cloudflare uses it to search its own blog and developer documentation. The main update concerns images. Previously, a picture was first described in text: objects were detected, a caption was generated, and only the caption was indexed. Search worked, but only on the details that made it into the caption.

How much it costs?

Billing starts on November 1, 2026, and Cloudflare promises an email reminder beforehand. You pay for three things: content uploads, storage and queries. Parsing, chunking, embedding with Workers AI models, keyword indexing and reranking are included in the price. There are no instance-hour charges and no monthly minimums.

What are the key details on Limitations and plans?

Video and audio are not supported yet; Cloudflare is only preparing a pipeline for them. The company is rewriting keyword search, because the current implementation has limits on large indexes. Separately, the team is working on simpler connection of AI Search to sites on Cloudflare, so AI agents can find and read content more easily.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Private AI Builders Community · Vibe Club

Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews

Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.

Join Vibe Club →Explore Hamidun Academy
What do you think?
Loading comments…