AWS Builder Center
Amazon Web Search on Search Arena

Amazon Web Search on Search Arena

We evaluated Amazon web search on Search Arena, both by training our own model optimized to search, as well as testing how Amazon web search performs when paired with other frontier models

In June, we launched  Web Search on Amazon Bedrock AgentCore  as a managed, MCP-compatible connector on AgentCore Gateway. Agents find it and call it like any other tool. To evaluate the performance of Amazon's web search, we tested it on Search Arena both when paired with our own trained search model, and when paired with leading frontier models.

What is Search Arena:
Search Arena is a crowdsourced evaluation platform that compares search-augmented Large Language Models (LLMs) using human preference votes. Users enter real-world prompts, and two anonymous, search-enabled AI models generate side-by-side responses. Users then vote on which answer they prefer. As a result, Search Arena goes beyond static factual tests (like BrowseComp or DeepSearchQA) to measure how well AI systems retrieve, synthesize, and cite information from the live web.


Our custom model trained to use Amazon's web search is now the #1 ranked system on the leaderboard, with an Arena score of 1,266.

Amazon Web Search Agent is a search agent we trained entirely on AWS primitives. We used SageMaker Training Jobs on SageMaker HyperPod EKS nodes to post-train an open-weight model, and it's built to use Amazon web search which is the same web search product that AWS customers can use today. Amazon web search is backed by a web index that Amazon operates, with tens of billions of documents refreshed continually. It also uses a knowledge graph for factual questions and pulls out only the relevant passages, so models get useful context instead of raw pages.

Beyond training our own model optimized to search, our Amazon web search performs competitively when paired with other frontier models too. We paired it with models from Anthropic and OpenAI to evaluate performance on Search Arena, although we did not publish these on the leaderboard itself:
  • Claude Opus 4.6 + Amazon Web Search: scored 1,273
  • GPT-5.6 Sol + Amazon Web Search: scored 1,252


Get started
  • Learn about  Amazon Bedrock Web Search : Connect your AI agents and applications to the web. Cited results you can trace, without your queries leaving your AWS environment by default.
  • Want Web Search Results?  Web Search on Amazon Bedrock AgentCore  is a managed, MCP-compatible connector on AgentCore Gateway that returns search results. Agents find it and call it like any other tool.
  • Want Web Answers with Citations?  Web Answer on Amazon Bedrock AgentCore  is a managed, MCP-compatible connector on AgentCore Gateway that returns web answers with citations. Agents find it and call it like any other tool.
Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.
Enjoyed reading this content? Let the author know!

Your likes, comments, shares, and saves help creators reach more builders.

Loading recommendations

Loading article