Human workloads
A team backs up once a night and serves media to visitors through apps. Objects are large, long-lived, and read by people.
Gigabytes dominate the invoice. Requests are rounding errors.
Agents use object storage as shared working memory: checkpoints, reports, and datasets passed between tools by URL. Frequent writes, repeated reads, and short object lifetimes change which provider costs least.
The same S3 API serves both, but the access shape barely overlaps. Billing follows the shape, not the technology.
A team backs up once a night and serves media to visitors through apps. Objects are large, long-lived, and read by people.
Gigabytes dominate the invoice. Requests are rounding errors.
An agent publishes checkpoints, diffs, reports, and datasets every few minutes. Readers are other agents, model APIs, and pipelines fetching plain URLs.
Operations dominate the invoice. Thousands of small writes per hour and repeated machine reads decide what you pay.
Ranked on a representative agent month: fifty gigabytes kept, one hundred fifty downloaded, five million writes, thirty million reads.
50 GB stored · 150 GB downloaded · 5M writes · 30M reads
| Provider | Monthly estimate |
|---|---|
| Backblaze B2 | $0.28Lowest |
| Scaleway Multi-AZ | $1.79+$1.51 vs lowest |
| IDrive e2 | $6.00+$5.72 vs lowest |
| AWS S3 Standard | $41.64+$41.37 vs lowest |
AWS S3 Standard is the compatibility baseline most agent frameworks default to. The full model, warnings, and sources are in the main calculator.
Three options cover most setups. The first two need no infrastructure from you.
Anthropic, OpenAI, and Gemini ship file storage inside their platforms, so agents can attach and share artifacts without any external bucket.
Services like EasySend or f2u expose MCP servers and no-auth upload APIs that turn local files into short-lived links on demand.
For production fleets, point agents at S3-compatible storage you control. That is where the pricing rules above decide your bill.