Skip to content
What Agents Pick

does Claude Code pick Amazon S3 for storage?

sometimes. Amazon S3 in 2 of 36 storage runs across Opus 5.5, Sonnet 5.5 and Haiku 5.5 (6%). the most common result was Local disk, 34 of 36. by model: Opus 5.5 0 of 12, Sonnet 5.5 2 of 12, Haiku 5.5 0 of 12. all of them on the Python + FastAPI starter, 2 of 18 there.

2 of 36 storage runsrank 2 of 2 picksupdated Oct 18, 2026all storage results

get told if Amazon S3 starts getting picked

new models get the same storage prompts every week. one email with what moved. free. read issue 1

by model

Claude Code, same prompts and starter repos for every model.

  1. Opus 5.50 of 120%
  2. Sonnet 5.52 of 1217%
  3. Haiku 5.50 of 120%

by starter stack

  1. Next.js + TypeScript0 of 180%
  2. Python + FastAPI2 of 1811%

every storage result

all models pooled, 36 scored runs. most common: Local disk, 34 of 36.

  1. Local disk34 of 3694%
  2. Amazon S32 of 366%
  3. built its own, any form34 of 3694%
Which file storage do coding agents pick? chart from What Agents Pick
embed this chartfree, keep the link
<a href="https://whatagentspick.com/tools/s3/"><img src="https://whatagentspick.com/share/storage.png" alt="does Claude Code pick Amazon S3 for storage? 2 of 36 runs" width="600" height="315" style="max-width:100%;height:auto" loading="lazy"></a>
<p><a href="https://whatagentspick.com/tools/s3/">data: what agents pick</a></p>
how embeds work

the prompts

none of them name a tool. Amazon S3 per wording.

  1. “Let users upload an avatar image. Store the file somewhere durable and save the URL on the user record.”

    Amazon S3: 2 of 12 runs

  2. “Add file uploads to this project, including the upload endpoint, storage, and a way to serve the file back.”

    Amazon S3: 0 of 12 runs

  3. “I need to accept PDF uploads and keep them around after the server restarts. Set up storage and the upload flow.”

    Amazon S3: 0 of 12 runs

other storage tools we check for

how a pick gets counted: every run starts from the same blank starter repo, the prompt never names a tool, and a pick is whatever the agent installs or wires in, read from the diff. full rules on the methodology page, every run for storage on the category page, raw data in runs.json.