AWS made Web Search generally available on Amazon (AMZN ) Bedrock on August 4, 2026, giving foundation models on its AI platform a built-in way to ground answers in current web knowledge without a third-party search provider. The server-side tool is enabled through a single parameter in an existing OpenAI-compatible API call, and responses come back with structured citations to the sources used, AWS said in its announcement.
The launch extends a capability AWS first shipped in June 2026, when Web Search went generally available on Bedrock AgentCore, the company’s agent infrastructure layer, at the AWS New York Summit. That version reached agents through the AgentCore Gateway and the Model Context Protocol. The new release moves the same underlying tool directly into Bedrock model inference, so any application calling a supported model can use it without standing up agent infrastructure first.
Until now, grounding a Bedrock model in live information meant selecting a search vendor, integrating its API, managing keys and orchestration, and putting the vendor through a security review. AWS is selling the removal of that work: no third-party vendors to onboard, no external APIs to orchestrate, and no additional vendor security reviews, per the announcement.
What the tool does and what it costs
Web Search runs on a web index that Amazon operates, spanning billions of documents and refreshed continually, combined with a built-in knowledge graph that anchors entities and the connections between them. For factual questions, the tool answers from the knowledge graph rather than leaving the model to infer an answer from extracted page text. Retrieval works through semantic snippet extraction, which pulls the passages relevant to the query instead of handing the model raw pages, keeping token spend on boilerplate down.
When the tool is enabled, Bedrock handles the full search lifecycle server-side: the model identifies a knowledge gap, Bedrock formulates the query, retrieves from the index and knowledge graph, injects snippets, source URLs, and titles into the context window, and returns a response with citation annotations marking the exact character span each source supports. If the first results are insufficient, the model can reformulate and search again within the same turn; if the results do not support an answer, the model says so rather than filling the gap from training data, according to the Web Search documentation.
The Amazon Bedrock pricing page lists Web Search at $12.00 per 1,000 queries in the three US Regions where it runs. That is a usage charge on top of model inference; AWS prices the AgentCore version of the tool separately.
The terms carry one obligation worth noting for teams building user-facing products: customers must retain and display the source citations and links from model outputs in anything surfaced to end users, and may not use the tool to bulk-extract content or build a competing index.
Where it runs and which models it supports
At launch, Web Search works only with OpenAI models served through Bedrock’s bedrock-mantle inference endpoint via the Responses API: openai.gpt-5.4, openai.gpt-5.5, and the three GPT-5.6 variants (Luna, Terra, and Sol), per the documentation. Anthropic, Meta, Mistral, and Amazon’s own Nova models on Bedrock are not on the supported list.
Availability is limited to three US Regions: us-east-1, us-east-2, and us-west-2, with queries processed in-Region. The documentation describes the tool as strictly regional: each Region operates its own search and fetch tier, and queries, fetches, index data, and results do not cross Regional boundaries.
The compliance posture
AWS is pitching the tool’s data handling as its enterprise differentiator. By default, retrieval is served entirely from the Amazon Bedrock web index and cache, a snapshot of web content hosted inside AWS, so request data does not leave the AWS boundary. Live-web retrieval, fetching content directly from the web at request time, is not yet served; the API parameter for it is already in place, and the documentation says any future change will require customers to explicitly grant a separate IAM permission before it takes effect.
Auditing runs through CloudTrail out of the box. Every search and fetch call is recorded with the calling identity, timestamp, and Region context, and denied requests log the specific condition key that caused the denial. By design, the trail does not record query text, returned URLs, or retrieved page content, giving compliance teams a record of who used the tool and when without exposing what users searched for.
From AgentCore to model inference
The AgentCore launch (announced at AWS New York Summit 2026) and the August 4 release both bring the Web Search tool to different layers of the Bedrock stack.
The distinction between the two versions is the integration point. The AgentCore tool is a connector target on a gateway, aimed at teams already building agents on that stack and invoked through MCP. The Bedrock version attaches to an ordinary inference call, which puts grounding within reach of a much larger set of existing applications, including ones with no agent architecture at all. Pricing for the two now sits on separate pages at separate rates.
With grounding reduced to a tools entry in the request body, the remaining work for adopters is permissioning: enabling the tool requires IAM actions for search and fetch alongside Bedrock inference access, and authentication to the bedrock-mantle endpoint uses a short-lived bearer token minted from existing AWS credentials rather than a separate API key. Live-web retrieval remains the dated-but-unscheduled item on the roadmap, gated behind a permission that is disabled by default.
