Amazon Bedrock web search alternatives in 2026

Give any Bedrock model SpecStream as a tool

The Bedrock web search tool runs only on OpenAI GPT models through bedrock-mantle. SpecStream works with every model the Converse API can give a tool to. Define one tool, call the SpecStream Search API when the model asks for it, and send back the answer and the ranked results as the toolResult. The model, the region and your AWS billing for inference stay as they are.

import os, boto3, requests

bedrock = boto3.client("bedrock-runtime", region_name="us-east-1")
MODEL_ID = os.environ["BEDROCK_MODEL_ID"]  # any Bedrock model that supports tool use

tool_config = {"tools": [{"toolSpec": {
    "name": "search_web",
    "description": "Search the live web. Returns the answer with its page, and the ranked results with title, URL and snippet.",
    "inputSchema": {"json": {"type": "object", "properties": {"q": {"type": "string"}}, "required": ["q"]}},
}}]}

def search(q):
    data = requests.get(
        "https://api.specstream.ai/v1/web-search",
        params={"q": q},
        headers={"Authorization": f"Bearer {os.environ['SPECSTREAM_API_KEY']}"},
    ).json()["data"]
    sources = [{"title": s["title"], "url": s["url"], "snippet": s["snippet"]} for s in data["sources"]]
    return {"answer": data["answer"], "sources": sources}

messages = [{"role": "user", "content": [{"text": "Who is the CEO of Mistral AI?"}]}]
while True:
    response = bedrock.converse(modelId=MODEL_ID, messages=messages, toolConfig=tool_config)
    message = response["output"]["message"]
    messages.append(message)
    if response["stopReason"] != "tool_use":
        break
    messages.append({"role": "user", "content": [
        {"toolResult": {
            "toolUseId": block["toolUse"]["toolUseId"],
            "content": [{"json": search(block["toolUse"]["input"]["q"])}],
        }}
        for block in message["content"] if "toolUse" in block
    ]})

print("".join(block.get("text", "") for block in message["content"]))

Add speed=fast to params for ranked results at $1 per 1,000 searches, and passages=true for the text of each page closest to the question.

Add SpecStream to your AgentCore gateway

An AgentCore Gateway takes SpecStream as an MCP server target next to Amazon's web-search connector. Store the SpecStream key in AgentCore Identity, then add the target. Agents on the gateway see specstream___search_web and specstream___read_pages with every other tool, and each search costs $0.005.

aws bedrock-agentcore-control create-gateway-target \
  --gateway-identifier <gateway ID> \
  --name specstream \
  --target-configuration '{"mcp":{"mcpServer":{"endpoint":"https://api.specstream.ai/mcp"}}}' \
  --credential-provider-configurations '[{"credentialProviderType":"API_KEY","credentialProvider":{"apiKeyCredentialProvider":{"providerArn":"<provider ARN>","credentialLocation":"HEADER","credentialParameterName":"Authorization","credentialPrefix":"Bearer"}}}]'

The AgentCore setup guide covers the key, the credential provider and the console path.

From $1 per 1,000, with no tokens on top

100,000 searches of 10 results cost $100 on SpecStream with speed=fast, $500 on the default search, $700 on AgentCore Web Search before gateway operations, and $1,200 on Bedrock Web Search before model inference. On SpecStream each reply carries price_usd, and the account's request log and usage totals show every call and its price.

The most accurate web search API in the world

The SpecStream Search API scores a 97.9 F-score on all 1,000 official SimpleQA Verified questions, with 978 correct. On WideSearch it scores 64.0 item F1 and 43.4 row F1 on all 200 official tasks. The scores and the method are at specstream.ai/benchmarks.

Results your code can read and keep

Bedrock Web Search hands your code the model's reply and its citations, and AWS says a citation can rest on a Search snippet even when the page was never fetched. A SpecStream search returns the answer with the page it was read on and its read time, rows with the facts the request asks for, and sources in rank order, each marked cited when a value comes from it. Every charged search stays in the account's history, and reading it costs nothing.

AgentCore's terms ask you to show the citation and link of each search result you use, and rule out storing results in bulk or building an index or database from them.

SpecStream

The live web data engine for teams and agents, and the best alternative to Amazon's web search tools. The Search API costs from $1 per 1,000 searches, works with every Bedrock model and any other, and scores 97.9 F-score on all 1,000 official SimpleQA Verified questions. The Extract API costs $0.80 per 1,000 pages. Deep Search returns every match as structured rows at 5¢, 15¢ or $1 a row.

OpenAI web search tool

A tool on the OpenAI API at $10 per 1,000 calls, plus search content tokens billed at the model's rates. SpecStream costs from $1 per 1,000 searches with no tokens, for any model. SpecStream as an OpenAI web search tool alternative.

Claude web search tool

A server tool on the Claude API at $10 per 1,000 searches, plus standard token costs for the content it brings back. SpecStream gives Claude on Bedrock or on the Claude API its search from $1 per 1,000. SpecStream as a Claude web search tool alternative.

Gemini grounding with Google Search

On Gemini 3 and newer, 5,000 search requests a month are free, then $14 per 1,000, with a default limit of 1,500 requests a day per project on the paid tier. On SpecStream your code decides how many searches run, from $1 per 1,000, from any model.

DeepSeek Search API

A raw JSON search API at $0.75 per 1,000 successful requests, with queries of 128 characters at most and up to 50 results from DeepSeek's own search engine. SpecStream takes questions of up to 2,000 characters and returns the answer with the page it was read on.