Cloudflare AI Search adds visual retrieval, with billing starting November 1

Cloudflare AI Search is generally available as of October 1, with new support for searching visual details in images and reading scanned PDFs. Developers can use it to build search over their own files or websites. Usage billing begins November 1, 2026, with monthly allowances remaining on every Workers plan.

The service accepts uploaded files, a connected Cloudflare R2 storage bucket or a website the customer owns. It prepares that material for retrieval, then finds relevant passages or images when a query arrives. Applications can return those results directly or ask a generation model to turn them into an answer.

Image search can look beyond the caption

Previously, AI Search made an image searchable by generating a description and searching that text. Details left out of the description were lost to retrieval. The new approach retains the caption and also processes the image itself, so a query can contain text, an image or a combination of the two.

Cloudflare supports native image retrieval through Qwen3-VL-Embedding. The model represents words and pictures as comparable lists of numbers, helping search find similar content. Our embedding explainer covers that mechanism. The configured model still matters: a text-only model receives a caption of an image query rather than its visual features.

A 10 MiB PDF limit comes with an OCR condition

Optical character recognition, or OCR, turns letters photographed on a page into text the service can search. It is off by default in AI Search, and enabling it triggers a full reindex of the material. Our scanned-PDF guide explains how to tell whether a document has readable text in the first place.

The current file documentation sets different size limits. MiB measures file size, not page count:

• Plain-text and code files can be up to 10 MiB each.
• PDFs can be up to 10 MiB when OCR is enabled.
• PDFs without OCR and other formats converted to Markdown remain limited to 4 MiB. Markdown is the text format the service uses for those converted documents.

What billing will cover from November 1

Each account gets five million monthly ingestion tokens, 10 GB-month of storage, 1,000 semantic searches and 1,000 full-text searches. Ingestion means preparing content for search, and tokens are the units into which the processed text is split. Images and text share the five-million-token allowance.

Cloudflare’s published overage rates are in US dollars:

• Ingestion costs $0.75 per million tokens, with image processing or scanned-document OCR adding $0.50 per million.
• Storage costs $2 per GB-month, combining the amount stored with how long it is kept.
• Semantic searches, which look for related meaning, cost $0.75 per 1,000. Full-text searches cost $0.10 per 1,000.

File preparation and result reranking with covered Workers AI models are included. Answer generation, AI query rewriting and external model providers are billed separately. Returning matching documents and writing an answer from them can therefore produce different bills.