Cloudflare AI Search has moved from beta to general availability. The bigger operational change arrives on November 1, when usage beyond the included allowance starts generating charges.
Cloudflare announced general availability on October 1. The managed retrieval service now uses hybrid search by default, combining semantic and full-text signals. It can ingest text, code, and documents, and it supports images through OCR and multimodal processing.
The monthly allowance is split across four meters.
The pricing documentation includes five million ingestion tokens, 10 GB-month of stored data, 1,000 semantic queries, and 1,000 full-text queries each month. Above those amounts, base ingestion costs $0.75 per million tokens, storage costs $2 per GB-month, semantic or hybrid queries cost $0.75 per 1,000, and full-text queries cost $0.10 per 1,000.
Image processing adds $0.50 per million ingestion tokens when that path applies. Workers AI calls used internally for embedding and reranking are included in AI Search and are not separately logged or billed. Generation and query rewriting can still add model costs outside the retrieval meters.
Worked example: a $78 retrieval month
Consider a workload that ingests 20 million base text tokens, stores an average of 25 GB-month, and runs 50,000 hybrid queries. Assume no image-processing surcharge and no separate generation cost.
- Ingestion: 15 million billable tokens after the five-million allowance, or $11.25.
- Storage: 15 billable GB-months after the 10 GB-month allowance, or $30.
- Hybrid queries: 49,000 billable queries after the 1,000-query allowance, or $36.75.
- Total: $78 for these listed AI Search components.
The example is not a quote for every application. Re-indexing, duplicate content, image OCR, query rewriting, and answer generation can change the total. Record each meter separately so a retrieval change is not mistaken for a model-price change.
File and OCR limits shape the index
Cloudflare documents a 10 MiB file limit for OCR, plain-text and code ingestion paths, and a 4 MiB limit for other supported files. Large PDFs may need deliberate splitting. Splitting by raw byte size can separate a heading from its content, so preserve document structure and retain a source URL or identifier on every chunk.
OCR makes image-heavy documents searchable, but it can misread tables, diagrams, and small text. Sample the extracted content before trusting retrieval quality. A screenshot that looks correct doesn’t mean the indexed text is correct.
GA checklist before November billing
- Export current ingestion, storage and query volumes by index.
- Remove duplicate or obsolete documents before a full re-index.
- Set alerts for token, storage, and query growth.
- Measure retrieval relevance separately from generated-answer quality.
- Retain source metadata so users can verify every retrieved passage.
- Estimate query rewriting and generation costs outside AI Search.
The cost-control method is the same one described in our AI agent spend guide: define meters, caps, and acceptance criteria before scale hides the waste. Cloudflare AI Search is easier to buy now that it is GA. It is still worth modeling the workload before November.
