Amazon Bedrock web search: AWS's $12-per-1,000-query toll

Amazon Bedrock web search is now a paid, first-party capability. AWS charges $12 per 1,000 queries to ground OpenAI's GPT models in current web knowledge. The tool reached general availability on August 4 in US East (N. Virginia), US East (Ohio), and US West (Oregon). Beyond the pricing, the launch is a strategy statement about who owns the retrieval layer of enterprise AI.

For teams already building on Bedrock, the change is immediate. Previously, adding web grounding meant taking on an outside search vendor, maintaining its credentials and payment account, building custom glue code, and getting that vendor cleared in security review. With Amazon Bedrock web search, that work collapses into a single web_search tool parameter on the OpenAI-compatible Responses API call. Bedrock handles the entire search lifecycle server-side, and one request returns the grounded answer with citations that include source URLs, page titles, and character offsets for inline footnotes.

I keep coming back to one number, because it reveals the strategy. $12 per 1,000 queries sits on top of model inference, and AWS charges by query rather than by token. Semantic snippet extraction trims context-window usage and cuts token spend relative to raw page retrieval, but that saving lands on the inference side of the bill; the search fee is a fixed toll per call. Work the arithmetic: an agent making five search calls per task generates $60 of grounding charges for every 1,000 tasks, before a single inference token is billed.

The obvious objection is that $12 per 1,000 queries is not cheap, and the Amazon-operated index is unproven against specialists on long-tail or niche queries. That misses who the buyer is. A regulated enterprise replacing a multi-vendor stack pays for the egress guarantee, the single audit trail, and the elimination of a vendor review cycle; marginal retrieval quality on obscure queries matters less than the compliance line items disappearing from the procurement sheet.

Why Amazon Bedrock web search changes the enterprise stack

Web Search is not a wrapper around a third-party index. Amazon operates the index behind it, covering tens of billions of documents refreshed continually, and maintains a separate knowledge graph that records entities and their relations. When a user asks for a fact, the answer is taken from that graph, so the model is not left to infer it from a page snippet. That is a direct strike at the hallucination problem. AWS frames the differentiators as multi-source grounding and single-parameter enablement, and the system draws on engineering from Alexa+, Amazon Quick, and Kiro, so the enterprise capability inherits consumer-grade search experience.

AWS describes the knowledge graph as a source of verified facts, a meaningful distinction from retrieval. For a question such as a company's founding date or a product specification, the tool returns an answer tied to the entity instead of a scraped page fragment the model must weigh on its own. That is the difference between searching and grounding, and it is the part of the design most likely to change how factual answers behave in production.

The announcement frames Web Search as a foundation-model grounding capability. The integration surface explains the launch lineup: the tool rides the OpenAI-compatible Responses API, so the web_search parameter slots into calls any model speaking that interface can make. GPT-5.4, GPT-5.5, and GPT-5.6 are simply the first models on Bedrock's next-generation inference engine, which leaves the door open for the same tool to serve other foundation models as they land on that engine.

The security design is where the pitch actually lands. Zero data egress is the default, processing stays in-region, and CloudTrail captures the audit trail without logging sensitive query text. Access runs through IAM and SigV4 with short-lived bearer tokens that expire within 12 hours. The permission model is explicit: bedrock-websearch:InvokeSearch and bedrock-websearch:InvokeFetch gate retrieval, with bedrock-websearch:ExternalWebAccess reserved for the live-web mode planned as a future update.

All of this is bad news for third-party search vendors that built businesses on being the grounding layer for enterprise LLMs. I read this as a deliberate disintermediation play: Amazon Bedrock web search replaces the multi-vendor stack in a single server-side call, with unified AWS billing, no vendor security review, and no data leaving AWS. Amazon has inserted itself into that position and is charging for it.

The OpenAI tension is part of the point

The most counterintuitive detail is the launch lineup: OpenAI models only, including GPT-5.4, GPT-5.5, and GPT-5.6 Sol, Terra, and Luna, served through Bedrock's next-generation inference engine at the bedrock-mantle endpoint. It is fair to ask why Amazon would ground its biggest model competitor. The answer is that hosting and grounding OpenAI workloads keeps them inside AWS infrastructure, security boundaries, and billing. Amazon does not need to win the model race to win the workload, and the data plane is how it collects.

What the announcement does not say is that the design attaches the platform to the retrieval path. Queries pass through AWS-operated infrastructure by definition, since zero data egress and in-region processing guarantee that, while CloudTrail gives the customer the audit record without exposing query text. That combination is why AWS can comfortably host a competitor's models: workloads, grounding, and audit trail all settle inside one bill.

The agent story makes the same point at a larger scale. Web Search went GA on Bedrock AgentCore in June at the AWS New York Summit, reachable through the AgentCore Gateway and the Model Context Protocol. This release moves the tool directly into model inference, so an application gets grounding without standing up agent infrastructure first. Agent loops issue many searches per task, and every one of those calls now flows through AWS as a metered line item.

For developers, the practical win is the citation structure. Bedrock returns source URLs, page titles, and character offsets in the same call that produces the grounded answer, so verification does not require a second retrieval pass or a separate fact-checking pipeline. Semantic snippet extraction also means the model sees only relevant passages, keeping token usage lower than raw page retrieval would allow.

One limitation keeps this from being a total sweep. Only indexed-web retrieval is live today; live-web retrieval is scheduled for a future update. The launch footprint is three US regions with no announced expansion timeline, so global teams are either waiting or routing workloads through those regions for now.

Why this matters

This launch turns grounding into metered infrastructure and sets a reference price for it: $12 per 1,000 queries inside Bedrock, on top of inference, with zero egress as the default. Amazon Bedrock web search is the benchmark every grounding vendor will now be measured against, and the warning that the separate grounding vendor is being squeezed out of the enterprise LLM stack. Teams choosing a search provider for their agents should price both the per-query fee and the compliance work it removes.

Sources

Introducing Web Search on Amazon Bedrock for foundation model grounding

Amazon Bedrock launches Web Search for OpenAI GPT models - AWS

Related Articles

• OpenAI GPT-5.5 and GPT-5.4 on Amazon Bedrock Reach General Availability for Enterprise Use
• GPT-5.6 on Amazon Bedrock: OpenAI's Sol, Terra, and Luna Models Arrive for Enterprise Deployments
• AWS Cuts AI Costs with Nova Model Distillation

✔Human Verified

Researched and cross-referenced against primary sources by the Bytevyte editorial team. This article was generated with the assistance of artificial intelligence and reviewed by the Bytevyte editorial team.