Server-side tools — web search, code execution, web fetch, computer use — bill differently from plain chat. Some are free beyond standard tokens, one has a flat per-call fee, one bills by execution time. Here's each one, straight.
| Tool | Extra charge |
|---|---|
| Web search | $10 / 1,000 searches |
| Web fetch | Free (tokens only) |
| Code execution (with search/fetch) | Free |
| Code execution (standalone) | $0.05/container-hour after 1,550 free hrs/mo |
| Computer use | Free (standard tokens + screenshots) |
| Bash / text editor | Free (standard tokens only) |
"Free" means no charge beyond ordinary input/output token pricing — every tool still adds tokens for its definition and results. Prices as of 2026-07 — confirm current numbers at platform.claude.com/docs/en/about-claude/pricing.
Every web search costs $10 per 1,000 searches ($0.01 each) on top of standard token costs for the query and the search-generated content that gets pulled into context — and that content can be substantial. Each search counts as one use regardless of how many results it returns, and a failed search isn't billed. Run 500 searches a day doing research or fact-checking and that's $5/day in fees alone, before a single token of the actual results is counted.
Web fetch has no extra charge at all — you pay standard token costs for whatever content gets pulled into the conversation (a typical 10 KB page runs ~2,500 tokens; a 500 KB PDF can run ~125,000 tokens, so the cost lives in what you fetch, not the fetch itself).
Code execution is free when it's used alongside web search or web fetch in the same request. Used on its own, it bills by container time instead of tokens: a 5-minute minimum per execution, 1,550 free hours per organization per month, then $0.05/hour per container after that. Note that if you attach files to a request, execution time is billed even if the tool never actually runs — files get preloaded onto the container regardless.
Computer use, bash, and the text editor tool all bill at standard token rates with no extra per-call fee — but each one adds fixed input tokens just for being present in a request, before you've used it once:
| Addition | Extra input tokens |
|---|---|
| Tool-use system prompt (any tool, Sonnet 5) | ~354–474 |
| Computer use system prompt overhead | 466–499 |
| Computer use tool definition | 735 |
| Text editor tool definition | 700 |
| Bash tool definition | 244–325 |
These stack: a computer-use agent with bash and text editor attached can add 2,000+ input tokens before a single screenshot or command result comes back — and each screenshot is priced as a normal image input (see vision pricing).
None of this is large per-call, but it's per-call — an agent looping through hundreds of tool calls in a long session accumulates this overhead on every single turn, not once per session.
Use the full calculator to price the token side of a tool-using request, then add web search at $0.01/call if your agent searches. See also: Claude Code cost for what an actual coding-agent session runs, and Batch API pricing for async workloads that don't need tools at all.