Cloudflare made AI Search generally available on October 1, 2026, with usage billing from November 1. It is a managed retrieval service: upload documents or point it at a site, and it handles the chunking, embedding, indexing and search an AI assistant needs to answer from your own content. This page compares it with Google Agent Search, Amazon Bedrock Managed Knowledge Base and OpenAI file search, priced from each vendor’s own pages.
- 01Cloudflare is cheapest per query$0.75 per 1,000 semantic queries against $1 to $4 at Google and AWS and $2.50 at OpenAI.
- 02Billing starts November 1Cloudflare AI Search is GA, but usage is not billed until November 1, 2026.
- 03Storage is measured differentlyRaw data at Google and AWS, chunks plus embeddings at OpenAI. The same corpus bills differently.
- 04Connectors split the fieldAWS connects to SharePoint, Confluence and Drive. OpenAI takes uploads only.
01 — The releaseWhat Cloudflare made generally available
AI Search started as an open beta called AutoRAG in April 2025. Cloudflare’s GA announcement reduces the bill to three meters: content ingested, data stored and queries run. The limits and pricing page sets them at $0.75 per million ingestion tokens, $2 per GB-month of storage, and $0.75 per 1,000 semantic, vector or hybrid queries, or $0.10 for plain full-text queries. Embedding and reranking with Cloudflare’s own Workers AI models are included rather than billed separately.
Each month is free up to 5 million ingestion tokens, 10 GB of storage, 1,000 semantic queries and 1,000 full-text queries. The GA release also raised the size limit for text files and PDFs read with OCR to 10 MiB, from 4 MiB. Hybrid search, which combines keyword and vector matching, is on by default for new instances.
Usage before November 1, 2026 is not billed. A team that wants a free month of real traffic has October; a team planning a budget should price from November, when the free allowances above become the only free usage.
02 — ContextWhat is being compared
Two of the four need a naming note, which matters when searching for their docs. Google renamed Vertex AI Search to Agent Search on April 22, 2026, and says the functionality is unchanged. AWS now sells two kinds of knowledge base. The table uses the newer one, which is priced per unit like Cloudflare’s.
Cloudflare AI Search
Ingestion, storage and query meters. Sources are uploads, R2 storage or your own site.
Google Agent Search
Per-query and per-GiB pricing, or a subscription. Google’s RAG Engine is a separate, assemble-it-yourself option with no list price of its own.
Bedrock Managed Knowledge Base
Per-GB and per-call pricing. The older customer-managed knowledge base bills the vector store you run instead.
OpenAI file search
Per-GB-day storage and per-call search, plus model tokens. Upload only, no connectors.
Microsoft’s Azure AI Search is left out of the tables. It bills by provisioned search units or compute hours rather than per query, so its prices do not line up with the four above.
03 — The dataStorage and query prices
Query prices are per 1,000. Storage is per month except at OpenAI, which bills per GB per day; at 30 days that is about $3 per GB-month. The storage figures do not measure the same thing: Google and AWS count the raw data you load, Cloudflare counts what sits in its index, and OpenAI counts the parsed chunks plus their embeddings.
| Service | Storage | Queries per 1,000 | Free each month |
|---|---|---|---|
| Cloudflare AI Search | $2.00 per GB-month | $0.75 semantic or hybrid; $0.10 full-text | 5M ingestion tokens, 10 GB, 1,000 + 1,000 queries |
| Google Agent Search (Standard) | About $5 per GiB-month | $1.50 | 10,000 queries, 10 GiB |
| Google Agent Search (Enterprise) | About $5 per GiB-month | $4.00, generative answers included | 10,000 queries, 10 GiB |
| Amazon Bedrock Managed Knowledge Base | $5.00 per GB-month | $1.00; agentic $4.00 plus $1.00 per underlying call | None stated |
| OpenAI file search | $0.10 per GB-day | $2.50 per 1,000 tool calls, plus model tokens | 1 GB of storage |
Google’s 10,000 free queries are described as a monthly trial allowance per account and exclude its advanced generative answers; the same page’s annual example deducts them only once, so treat the Google totals below as a best case. AWS publishes no free tier for the managed knowledge base. Google also offers a subscription model with a minimum commitment of 1,000 queries a minute and 50 GB of storage, aimed at steady high volume.
04 — Worked exampleOne workload, four bills
Take 10 GB of documents and 100,000 retrieval queries a month. Apply each list price and free allowance, and leave out ingestion, generation and model tokens. This is our arithmetic, not a vendor quote. Cloudflare: storage fits the free 10 GB, and 99,000 billable queries cost $74.25. Google Standard: storage is free, and 90,000 billable queries cost $135. AWS: $50 of storage plus $100 of queries is $150. OpenAI: 9 GB beyond the free one for 30 days is $27, and 100,000 calls are $250, a total of $277 before model tokens.
Monthly cost for 10 GB and 100,000 queries, US dollars, lower is cheaper
Digital Applied arithmetic from list prices read October 3, 2026. Excludes ingestion, generation and model tokens; OpenAI also bills model tokens on top. Storage bases differ by vendor.The ranking flips with the shape of the workload. A large corpus queried rarely favours the cheapest storage; a small corpus queried constantly favours the cheapest query. AWS’s own pricing page gives a larger example, 50 GB and 100,000 standard retrievals at $350 a month, or $850 with agentic retrieval, in which a managed model plans the retrieval and each underlying search is billed as well.
05 — The dataIngestion, files and sources
| Service | Max file | Indexing costs | Data sources |
|---|---|---|---|
| Cloudflare AI Search | 10 MiB text and OCR PDFs; 4 MiB others | $0.75 per 1M tokens (+$0.50 for images); Workers AI embedding and reranking included | Upload, R2 buckets, your own site on the same Cloudflare account |
| Google Agent Search | 200 MB | Included in storage; Layout Parser $10 per 1,000 pages; Ranking API $1 per 1,000 | Websites, Cloud Storage, BigQuery and other Google databases, Workspace; third-party connectors no longer supported |
| Bedrock Managed Knowledge Base | 30 MB of extracted text | Managed parsing, embeddings and reranker at $0 | S3, SharePoint, Confluence, Google Drive, OneDrive, web crawler, custom |
| OpenAI file search | 512 MB; 5M tokens per file | Not priced separately | File upload only |
Connectors are where the products differ most. Agent Search has dropped outside data sources, according to Google’s docs, and those connectors now live in its Gemini Enterprise product. AWS’s managed knowledge base reads from SharePoint, Confluence, Google Drive and OneDrive and can filter results by each document’s permissions, which its documentation says the customer-managed version cannot. Cloudflare crawls only websites on the same Cloudflare account. For a team turning its own content into a source, our method for turning a blog archive into a knowledge base covers the preparation that comes before any of these services.
06 — The dataScale, rate and data limits
| Service | Scale | Rate | Data controls |
|---|---|---|---|
| Cloudflare AI Search | 1M files per instance (500K with hybrid); 5,000 instances on paid plans | Not published | No index residency statement; EU-jurisdiction R2 buckets accepted as a source |
| Google Agent Search | 10M documents per project per location | 300 searches a minute per project and location | Data at rest in us or eu multi-regions, or global |
| Bedrock Managed Knowledge Base | 10 TB per knowledge base; 200 data sources each | 100 retrievals a second per account; 600 a minute per knowledge base | Eight regions including GovCloud; document-level permission filtering |
| OpenAI file search | Not published on the pages read | 100 to 1,000 a minute by usage tier | Data residency supported; vector stores not eligible for zero retention |
OpenAI’s data controls page lists vector stores as not eligible for zero data retention: stored files stay until they are deleted. Its retrieval guide documents an expiry setting that deletes a store and stops its charges. Whichever service you choose, deletion is part of the design, as our guide to where deleted agent data survives explains.
07 — Practical implicationsWhich one to use
Retrieval quality is not on this page, because none of the vendors publishes a comparable measure and we did not run one. Test with 50 real questions from your users before committing, and check that each answer cites the right document. For agents that also need the open web, the sibling comparison of web search APIs for AI agents prices that side. Our AI transformation work covers the evaluation, the build and the cost controls.
08 — MethodMethod and as-of date
A comparison of published prices and documented limits. Nothing on this page was benchmarked by Digital Applied.
- What was collected
- Storage, query, ingestion and add-on prices, free allowances, file size limits, scale and rate quotas, data sources and data controls for four managed retrieval services, with Google’s Standard and Enterprise editions as separate price rows.
- Sources
- Each vendor’s own pricing pages, quota pages, product documentation and announcements: Cloudflare’s GA post and AI Search docs, Google Cloud’s Agent Search pricing, quota and release notes, AWS’s Bedrock pricing, What’s New post and user guide, and OpenAI’s pricing and retrieval guides.
- As-of date
- October 3, 2026. Cloudflare’s pages were updated October 1 and Google’s release notes and AWS’s pricing page September 30. Google’s pricing and quota pages, AWS’s documentation and OpenAI’s pages show no content date, and OpenAI’s prices could not be confirmed against an archived copy from before October 2.
- Units
- US dollars. Queries per 1,000. Google prices storage per GiB-hour; the GiB-month figure is the page’s own rounded example. OpenAI’s per-day storage is converted at 30 days.
- Exclusions
- Azure AI Search, which bills by capacity rather than per query; Google RAG Engine, which has no list price of its own; the customer-managed Bedrock knowledge base, whose cost is the vector store you provision; enterprise and committed-use discounts.
- Limitations
- No retrieval quality, latency or indexing speed was measured. Cloudflare publishes no query rate limit or index residency statement. OpenAI’s file and store counts were not on the pages read.
- Refresh
- Re-read every pricing page monthly, and after November 1 when Cloudflare billing begins. Correct any figure in place with a dated note.
Price your corpus and query volume before choosing
Measure two numbers first: how many gigabytes you will index and how many queries a month you expect. Run them through the four price rows, rule out any service that cannot reach where your documents live, and test the survivors on real questions.