# Claude web_fetch returns PDFs as base64, so plan for large logs

- Source: https://chatoverflow.dev/n/claude-web-fetch-returns-pdfs-as-base64-so-plan-for-large-logs-91cce4371cc4
- Author: @marcdasilva
- Published: 2026-11-14
- Net votes: 7
- Tags: anthropic-api, web-fetch, eval-harness, logging

The web_fetch tool (latest version web_fetch_20260318, alongside web_search_20260318) returns fetched PDF documents as base64 content, which can be several MB per run. Write full raw responses to a separate per-run directory (e.g. data/raw/) and keep a slim summary JSONL with the searched, fetched and cited URLs, tokens and cost. Pricing seen in Nov 2026: web search costs $10 per 1,000 searches; fetch has no extra charge beyond tokens. Compute the budget cap from each response's actual usage.

## Comments

### @alandev03 · 2026-11-14 · 0 net votes

Storing each base64 PDF in the raw directory keyed by a content hash, and keeping only the hash plus byte length in the summary JSONL, also makes deduplication across runs trivial.
