Sulat.com / Writing on AIFriday, November 20, 2026
Artificial Intelligence/20 November 2026

Amazon’s web search agent takes #1 on Search Arena with a 1,265 score

Amazon-Web-Search-Agent ranks first of 43 models on Search Arena at 1,265, five points ahead of second place. AWS trained it on SageMaker and built it to use Amazon's own web index.

Glossy 3D Amazon logo above a number 1 podium with a magnifying glass and search bar

At a glance

  • Amazon-Web-Search-Agent ranks first of 43 models on Search Arena at 1,265, plus or minus 15.
  • Qwen3.8-Max is second at 1,260, so the five-point lead sits inside the margin.
  • AWS post-trained an open-weight model on SageMaker HyperPod and hasn't named the base model.
  • Amazon web search costs US$7 per 1,000 queries through its AgentCore connector and API.

Amazon’s Amazon-Web-Search-Agent has taken first place on Search Arena, the leaderboard where people vote between two AI answers built with live web search. Arena posted the result on X, saying the model, coupled with AgentCore Web Search, “takes the #1 spot on Search Arena.” Arena’s model page lists a score of 1,265 across 1,574 votes, and AWS published its own write-up the same day.

What Search Arena measures

Users type a real question and get two anonymous answers from search-enabled models, then vote for the one they prefer. AWS contrasts that with static factual tests such as BrowseComp and DeepSearchQA. Arena turns the votes into scores, and the board lists 43 models.

The score and the margin

Amazon-Web-Search-Agent sits first of 43 at 1,265, with a margin of plus or minus 15. Qwen3.8-Max is second at 1,260, and Claude Opus 4.6 Search scores 1,252. The five-point lead over second place is smaller than that margin, and Arena’s rank spread for the model runs from 1 to 5. AWS’s post gives the score as 1,266.

Four cards: first of 43 models, score 1,265, 1,574 votes, rank spread 1 to 5
Amazon leads at 1,265, but the margin on that score is plus or minus 15. Source: Arena.

Vote counts vary widely across the board. On Arena’s 13 October snapshot, Claude Opus 4.6 Search had 173,191 votes and Qwen3.8-Max had 1,870. Amazon’s agent has 1,574.

How Amazon built the agent

AWS says it post-trained an open-weight model using SageMaker Training Jobs on SageMaker HyperPod EKS nodes, so everything ran on its own cloud. Amazon hasn’t named the base model, and Arena lists the agent’s licence as proprietary.

The agent is built to call Amazon’s web search, the same product AWS customers can buy today. That search runs on a web index Amazon operates, with tens of billions of documents refreshed continually. For factual questions it also draws on a knowledge graph, and it hands the model only the relevant passages instead of whole pages.

Four steps: open-weight model, post-trained on AWS, wired to Amazon web search, judged by human votes
Amazon trained the agent on its own cloud and pointed it at its own search index. Source: AWS.

Amazon’s search with other models

AWS also paired Amazon web search with frontier models and didn’t publish those runs on the leaderboard. Claude Opus 4.6 with Amazon web search scored 1,273, against 1,252 for Claude Opus 4.6 Search on the board. AWS ran that test itself, so the board doesn’t carry the 1,273.

Number line: Claude Opus 4.6 Search 1,252, Qwen3.8-Max 1,260, Amazon agent 1,265, Opus 4.6 plus Amazon search 1,273
AWS’s own run of Claude Opus 4.6 with Amazon search scored 1,273, a result the leaderboard doesn’t list. Source: AWS.

What you can buy today

Web Search on Amazon Bedrock AgentCore reached general availability in June as an MCP connector on AgentCore Gateway. On 10 November, AWS added the InvokeWebSearch API, which returns ranked results from a single call. Each result carries a text extract and, when available, the page title, URL and publication time. Requests can filter by domain and publication time, and IAM credentials authorise them.

Both routes cost US$7 per 1,000 queries. The API is available in US East (N. Virginia and Ohio), US West (Oregon), Asia Pacific (Sydney and Tokyo) and Europe (Ireland). AWS says the same search infrastructure runs Alexa+, Kiro and Amazon Quick. Its write-up points readers to those search products and doesn’t announce a way to call Amazon-Web-Search-Agent itself.

Hero illustration features the Amazon logo, which remains the trademark of its owner.


Discover more from Sulat AI & Tech News

Subscribe to get the latest posts sent to your email.