LATEST
AI Tools

AWS AgentCore Web Search API Is Live: $7 per 1,000 Searches

Updated November 16, 2026
14 min read
Editor's Note: This article was human-reviewed for accuracy
Close-up of a magnifying glass on a blue surface, symbolising web search for AI agents

Quick Answer & Key Takeaways

  • • AWS has made AgentCore web search callable with a single signed API call, InvokeWebSearch, at $7 per 1,000 queries, with no gateway to set up.
  • • The API returns ranked results from Amazon's own web index, and the per-query price is the same on the gateway route — the saving is the plumbing, not the search.
  • • The catch is governance: the administrator domain lists that the gateway enforced do not carry over, so the filtering now lives in your own code.

AI agents are only as current as their last web search. Until this month, giving an agent built on Amazon Bedrock AgentCore that ability meant creating a gateway, adding a web search target, and wiring up an MCP client, even when your code already knew when to search. AWS has removed that step. The Web Search API, generally available since November 10, 2026, lets any application call InvokeWebSearch directly: one signed call, no gateway, no separate API key, and the same $7 per 1,000 queries the gateway route charges. The price is unchanged. What changes is who holds the controls, and that is where the launch gets interesting. This article covers what the API returns, what it costs at realistic volumes, how it compares with Exa, Brave and Tavily, and the governance trade-off to settle before switching.

What AWS actually launched

AWS made the Web Search API generally available on November 10, 2026, and published its launch post the next day. InvokeWebSearch is an operation on the bedrock-agentcore data plane: your code calls an endpoint such as bedrock-agentcore.us-east-1.amazonaws.com, and the request is signed with AWS Signature Version 4 using the credentials the application already has. There is no API key to store or rotate, and, in AWS's own words, there are "no resources to create" before the first call.

The search runs against a web index that Amazon operates itself, the same infrastructure behind Alexa+, Kiro and Quick, according to AWS's announcement. The index spans tens of billions of documents and reflects new content within minutes, both AWS's own descriptions. Queries are served inside AWS and are not sent to a third-party search engine. At launch the API runs in six regions: US East (N. Virginia), US East (Ohio), US West (Oregon), Europe (Ireland), Asia Pacific (Tokyo) and Asia Pacific (Sydney). The same search had been available since June 16, 2026 only as a gateway connector, so this is a new front door to an existing service, not a new index. AWS still recommends the gateway for most agentic workloads.

What a call returns

Two fields are required: webSearchId, which must be web-search.v1, and searchQuery, from 1 to 200 characters. The optional fields are maxResults (1 to 25, default 25) and a filters object holding a domain filter with include and exclude lists of 1 to 100 domains each, plus a published-date filter with publishedAfter and publishedBefore timestamps. Domain entries are bare names such as example.net, and a root domain also matches its subdomains. The body can contain only those fields; anything else fails with a ValidationException.

The response is a results array in ranked order, with up to maxResults entries and no second page. Every result has a text extract from the page. The title, URL and publication time appear only when known, so citation code has to handle results with no URL. The API involves no model, does not synthesise an answer, and does not fetch full pages. A finished answer with citations is a different AWS product, covered below.

What it costs, with the fine print

The price is $7 per 1,000 queries, or $0.007 per query, billed per query submitted, with no minimum and no upfront commitment. In local currency that is about £5.30 or C$9.99 per 1,000 queries. Conversions use November 16, 2026 mid-market rates of 1 USD = 0.7576 GBP and 1.4277 CAD, and are approximate. There is no free tier for Web Search itself; new AWS customers receive up to $200 in general Free Tier credits, which is account credit, not a search allowance.

Queries per monthWeb Search charge (USD)GBP approx.CAD approx.
1,000$7.00£5.30C$9.99
10,000$70.00£53.03C$99.94
100,000$700.00£530.32C$999.39
200,000$1,400.00£1,060.64C$1,998.78

The 200,000-query row is AWS's own pricing-page example: a market research agent at that volume pays $1,400 for search, plus $3.00 in gateway invocation charges on the old route, for $1,403 in total. That $3 is the point: gateway invocations cost $0.005 per 1,000, so the gateway portion of a web-search bill is a rounding error. Moving to the direct API saves almost nothing in money; it saves the gateway itself.

One number matters more for budgeting than the per-query rate: the default quota is 10 requests per second per account per region, adjustable through AWS Support. An account running flat out at that cap for a 30-day month in one region would submit 25.92 million queries, or $181,440 (about £137,459 or C$259,042). A loop at the cap spends about $252 an hour. Throttled calls come back as ThrottlingException (HTTP 429); AWS's guidance is retry with exponential backoff. A burst is the realistic risk: 50 tasks starting at once, each firing five searches in parallel, is 250 requests against a cap of 10.

How the price compares

AWS sits in the middle of the search-API market on list price. The table uses each vendor's published list rate per 1,000 searches, checked November 16, 2026.

ProviderPrice per 1,000 searchesFree allowanceWhat you get
AWS AgentCore Web Search API$7.00 (£5.30 / C$9.99)None; up to $200 new-account AWS creditRanked extracts, up to 25 results
Exa Lite$4.00 (£3.03 / C$5.71)$10 credits a monthUp to 10 results; Auto $7, Deep $12
Brave Search API$5.00 (£3.79 / C$7.14)$5 credit a monthSnippets; LLM Context included
Tavily (basic, pay as you go)$8.00 (£6.06 / C$11.42)1,000 credits a monthText included; advanced $16
Google Search grounding$14.00 (£10.61 / C$19.99)5,000 free queries a monthGrounded answers, not raw results
Amazon Bedrock Web Search tool$12.00 (£9.09 / C$17.13)None publishedModel-run grounding in the Responses API

Two caveats before anyone switches on price alone. First, the per-query fee is usually the small number: the tokens your model reads from the returned extracts cost more than the search that produced them, so a cheaper search API that returns longer extracts can cost more overall. Second, the products are not interchangeable: Exa and Tavily include content extraction, Brave sells an independent index at a flat rate, and Google's grounding returns answers rather than ranked results. AWS's edge is operational, not price. The bill lands on your existing AWS account, authorisation is IAM, and there is no second vendor or second key to manage. For teams already running agents on AWS, that is worth more than the $2 per 1,000 Brave would save. Sourcing note: the Google and Bedrock-tool figures come from Metarix's November 11, 2026 reading of the vendors' own pricing pages, and the Tavily and Brave rates are their published list prices as tracked by independent pricing comparisons.

API or gateway connector: the differences that matter

PropertyInvokeWebSearch APIWeb Search Tool connector (MCP via Gateway)
How it is calledSigned HTTPS call to the AgentCore data planeMCP tools/call to a gateway endpoint
SetupOne IAM permissionGateway + target + service role
AuthorisationIAM onlyGateway inbound auth (IAM, JWT)
Default result countUp to 25Up to 10
Date boundspublishedAfter, publishedBeforefrom, to
Date field in resultspublishedAtpublishedDate
Administrator domain listsNone; filters are per requestAdmin lists, hidden from the agent
RegionsSix commercial regionsSame six, plus GovCloud (US-West)
Price$7 per 1,000 queries$7 per 1,000 plus gateway invocation fees

Three porting traps deserve attention. A date-only publishedBefore on the API is sent as midnight UTC at the start of that day, so asking for everything published up to October 31 silently drops almost all of October 31; AWS's guidance is to use the last second of the day, 2026-10-31T23:59:59Z. The response field changes from publishedDate to publishedAt, so a parser reading the old key sees every result as undated. And the default result count rises from 10 to 25, which feeds more text into a model's context on any call that omits maxResults. Treat a move between routes as a migration with tests, not a find-and-replace.

Calling it in four steps

First, pick one of the six supported regions. Second, grant the single IAM permission on the AWS-owned resource (the ARN's account field is the literal string aws):

{
  "Version": "2012-10-17",
  "Statement": [{
    "Effect": "Allow",
    "Action": "bedrock-agentcore:InvokeWebSearch",
    "Resource": "arn:aws:bedrock-agentcore:us-east-1:aws:tool/web-search.v1"
  }]
}

Third, update the SDK: Boto3 1.43.127 or later, or the equivalent in the AWS CLI. Fourth, call it. AWS's own example, lightly trimmed, shows the shape of a filtered call:

import boto3
client = boto3.client("bedrock-agentcore", region_name="us-east-1")
response = client.invoke_web_search(
    webSearchId="web-search.v1",
    searchQuery="Amazon Bedrock AgentCore release announcements",
    maxResults=5,
    filters={"domainFilter": {"include": ["example.net"]}},
)
for result in response["results"]:
    print(result.get("title"), result.get("url"))
    print(result["text"])

Note the .get() on title and URL and direct indexing only on text: that matches the response contract, since only text is guaranteed. AWS also published a worked example that builds a daily dependency watcher with LangGraph on top of the API; one run makes at most 12 searches.

The governance catch

On the gateway route, an administrator sets target-level domain lists that the calling agent cannot see or override, and request-level filters can narrow but never expand that scope. The API has no equivalent: each request carries its own domain filter, and nothing in the public documentation constrains what that filter can contain. There is no IAM condition key for search domains in the service reference, so a policy cannot limit the filter's contents.

That matters for agents that take instructions from untrusted text: an agent steered by a prompt injection could drop a domain filter or swap it for one that suits the attacker. On the gateway route the administrator's lists still apply; on the direct route the filter is whatever the code sends. The fix: never let the model construct the filters object. Build it in code from configuration, and treat any model-proposed domain as a request to narrow within that set, never to widen it.

The same permission gates both routes, which creates a second trap. The gateway's service role needs bedrock-agentcore:InvokeWebSearch on the same AWS-owned ARN the new API uses, so an organisation cannot allow the gateway route and forbid the direct route by denying the action. A blanket deny would block the gateway's role too. The separation has to be made by principal: grant the action to gateway service roles and to the specific workloads approved for direct calls, and to nobody else. Teams that handed out broad bedrock-agentcore:* permissions during AgentCore experiments should check whether those roles can now search the web directly, without any gateway domain list in the path.

Audit has a gap too. Calls are logged as CloudTrail data events, which are off by default and need a trail or event data store with an advanced event selector for the AWS::BedrockAgentCore::WebSearch resource type. What gets logged is who searched, when and how often: the recorded parameters hold only webSearchId and, if set, maxResults. CloudTrail does not record the search text or the filters, and AWS's product FAQ says that is by design. Throttled, denied and invalid requests are logged; unsigned, oversized or endpoint-policy-blocked requests are not. Teams that need to reconstruct what their agents did must log queries themselves, at the application layer. Calls can stay inside a VPC through a PrivateLink interface endpoint, and an endpoint policy can restrict callers by IAM principal: a second gate, but one that governs the network path, not the query.

What AWS has not shown yet

Most of what is known about this service comes from AWS, and several questions that matter for production use have no public answer. Freshness is the first. AWS says the index reflects new content within minutes, but the one independent measurement available points the other way for ordinary pages: in August 2026, a Classmethod engineer tested the connector in Tokyo and found 84 of 99 high-traffic articles (85%), but 0 of 1 article published within 24 hours, 20% to 33% at two to four days old, and 67% to 83% at four to nine days old. The sample is small, and the test measured whether a specific page appeared rather than whether the index had crawled it. Anyone building alerting or news monitoring on this API should test their own target sources first.

AWS publishes no latency figures or relevance benchmarks for the API. The product page carries a customer statement attributed to Prox: more than 5,800 research sessions and nearly 75,000 successful search calls in July, "with 99.8% completing in under one second" — but that predates the API and describes the connector route. The index-size wording also varies: the June technical post says "tens of billions of documents" while the product page says "billions of documents". Readers weighing a bigger AWS agent build may also want our earlier look at AWS Infinity, the open-source agent runtime AWS launched in October, and the Google Data Agent Kit, a free MCP toolkit for data-cloud services, for the other cloud's approach.

Which AWS route should you pick?

AWS now offers four ways to put web results in front of a model, and the choice is mostly about who should decide to search. If your code should decide, use the API. If an MCP-framework agent should decide and a platform team wants one place to set policy, use the gateway connector. If you want an answer rather than evidence, use Web Answer: a sibling connector returning a finished, cited markdown answer at $10 per 1,000 tool calls (about £7.58 or C$14.28), generally available in early November 2026 in US East (N. Virginia) only. If you already call a supported OpenAI model through the Responses API on Bedrock and want the model to run the whole loop, use the Bedrock Web Search tool at $12 per 1,000 calls. The API is the lowest-priced of the four AWS routes, and also the one that hands you the most responsibility. DeepSeek turned its own chat app into a priced API the same week, with per-use tool fees of its own.

Common mistakes to avoid

  • Treating $7 as the whole bill. The token cost of feeding extracts into a model is usually larger than the search fee. Budget the full task, not the query.
  • Copying the MCP field names. The API uses searchQuery, publishedAfter/publishedBefore and publishedAt; the connector uses query, from/to and publishedDate. A mechanical rename breaks silently.
  • Using a date-only publishedBefore. It resolves to midnight UTC at the start of that day and drops the day you meant to include. Use the last second of the day.
  • Letting the model build the filters. Domain and date filters are authority, not suggestions. Build them in code from configuration.
  • Assuming CloudTrail shows what was searched. It shows who, when and how often, never the query text or the filters. Log queries at the application layer if you need that record.
  • Calling without backoff. The 10-requests-per-second cap is per account per region, and a burst of parallel agent tasks will hit it. Retry with exponential backoff on 429s.

How much does the AgentCore Web Search API cost?

$7 per 1,000 queries, or $0.007 per query, billed per query submitted with no minimum or upfront commitment. That is about £5.30 or C$9.99 per 1,000 at November 16, 2026 mid-market rates. There is no free tier for Web Search itself, though new AWS accounts receive up to $200 in general Free Tier credits.

Do I need an AgentCore Gateway to use it?

No. The API is a direct operation on the AgentCore data plane and needs nothing provisioned beyond one IAM permission. The gateway route still exists, and AWS recommends it for most agentic workloads, especially where a platform team wants administrator-set domain policy.

Which regions support InvokeWebSearch?

Six commercial regions at launch: US East (N. Virginia), US East (Ohio), US West (Oregon), Europe (Ireland), Asia Pacific (Tokyo) and Asia Pacific (Sydney). The gateway connector is also available in GovCloud (US-West); the direct API is not.

Does it return full page content?

No. Each result carries a text extract from the page, plus the title, URL and publication time when known. It does not fetch full pages and does not synthesise an answer. For a finished, cited answer, AWS's Web Answer connector costs $10 per 1,000 tool calls.

Is there a free tier?

Not for Web Search. The pricing page lists no free allowance; the only offset is the up-to-$200 general Free Tier credit for new AWS customers. Rivals are more generous: Exa gives $10 of credits a month, Brave renews $5 of credit a month, and Tavily includes 1,000 credits a month.

AWS has priced the new API exactly where the old route was, which tells you what this launch is: not a price cut, but a removal of plumbing. The teams that gain most are the ones with their own agent loops and their own filters built in code. The teams that should stay on the gateway are the ones relying on administrator domain policy, or on CloudTrail to show what their agents searched for. If that is you, keep the gateway, and audit which roles hold the InvokeWebSearch permission before the direct route finds them.

Sources

Currency conversions use November 16, 2026 mid-market rates (1 USD = 0.7576 GBP, 1.4277 CAD), verified at publication, and are approximate.

14 min read
4 views
edit_note Written by

Ratul Ahamed

Founder & Editor-in-Chief

Ratul Ahamed is the founder and editor of AI Core Updates, an independent site covering AI tools, models and pricing. He is a web developer based in Bangladesh and also runs All In One Tools, a free collection of browser-based utilities. Contact: [email protected]

Read all stories by Ratul Ahamed arrow_forward
verified_user Human-Reviewed Editorial

forum Comments (0)

Leave a Reply

chat_bubble_outline

No comments yet. Be the first to join the conversation!

Weekly AI Digest

Get our weekly AI tools digest

One email a week analyzing new AI tools, model updates, and practical workflows. Tested and summarized with zero fluff. Unsubscribe any time.

bolt
Hands-on AI Tool Reviews
Pricing math, limits, and alternatives.
verified_user
Strict Anti-Spam Policy
1-click unsubscribe at any time.
mark_email_unread

Subscribe to Digest

Sent weekly • No spam

person_outline
mail_outline

No spam ever. You can unsubscribe with a single click at any time.

cookie

Cookie Settings

We use cookies to enhance your browsing experience, serve personalized content, and analyze our traffic. By clicking "Accept All", you consent to our use of cookies. Learn more