Cloudflare postpones AI Search billing and changes its default sync interval Cloudflare postpones AI Search billing and changes its default sync interval

The grace period has boundaries

The pricing documentation separates AI Search’s included infrastructure from charges elsewhere in an application. Workers AI embedding and reranking do not appear as separate Workers AI charges or in AI Gateway logs. Generation and query rewriting continue through the account and gateway instead. There is also a historical storage detail: website pages now use built-in storage, but objects left in the older dedicated R2 bucket can still contribute to storage usage. AI Search no longer writes to that bucket.

What the December bill covers

Each account receives five million ingestion tokens, 10 gigabyte-months of storage, 1,000 semantic queries and 1,000 full-text queries monthly. Beyond those allowances, base ingestion costs $0.75 per million tokens and storage costs $2 per gigabyte-month. Semantic, vector and hybrid queries cost $0.75 per thousand; full-text queries cost $0.10 per thousand. Image processing adds $0.50 per million tokens. Ingestion is measured after parsing and chunking, so text repeated across overlapping chunks counts each time. Embedding and reranking are included, along with vector indexing and website crawling. Generation, query rewriting and external model providers remain separately billed through the applicable service.

Scheduled refresh has alternatives

Website and R2 sources use synchronization jobs to process additions, changes and deletions. A deployment pipeline or content-management webhook can trigger a job instead of waiting for the scheduled run, with triggers permitted every 30 seconds. Dashboard controls also support cancelling jobs, pausing indexing and re-indexing individual files. Processing time depends on the number, size and format of files, plus model latency. An interval therefore describes the schedule, not a promise that every document becomes searchable within that period. After 31 days without search requests, external-source instances pause scheduled jobs. They remain searchable. After renewed search or chat traffic, activity checks resume synchronization; manual resumption is also available.

Direct uploads follow a different path

AI Search already includes storage and a vector index without requiring a separate R2 or Vectorize setup. Files sent through the Items API or dashboard are indexed immediately rather than waiting for scheduled synchronization. An instance can combine that upload path with a website or R2 source. For example, shared website documentation can coexist with additional uploaded files in the same instance. This distinction matters for applications that mix published reference material with documents supplied by their users: the two inputs do not share one refresh schedule.