Nearly everything you pay for is tokens in and tokens out. A handful of server-side tools bill on top of that, and it's worth knowing which before you build a feature around one.

Web search

Charged per search, at $10 per thousand searches, plus the tokens for whatever comes back. Results count as input tokens for the rest of the conversation, not just the turn they arrived in, so a search-heavy chat carries them forward.

A failed search isn't billed. Each search counts once regardless of how many results it returns.

If a feature searches on every message, that's a cent per message before tokens. Fine at small scale, a real line item at volume.

Web fetch

No additional charge. You pay tokens for the fetched content and nothing else.

That makes fetching much cheaper than searching, which is worth knowing when you're designing a research feature. Rough sizes: an average web page is around 2,500 tokens, a large documentation page around 25,000, a research paper PDF around 125,000.

That last figure is the one to watch. Fetching a few PDFs can quietly cost more than a day of ordinary use, which is why there's a parameter to cap how much content a fetch can pull in. Set it.

Code execution

Free when used alongside web search or web fetch. Used on its own it's billed by container time, with a generous free monthly allowance and a small hourly rate after that. Execution time has a five-minute minimum, and if you attach files, the clock runs whether or not the tool is called, because the files get loaded onto the container regardless.

Managed agent sessions

If you use Anthropic's hosted agent sessions, runtime bills at a per-hour rate alongside tokens, counted only while the session is actually running rather than idle. It replaces container billing rather than adding to it.

Your own tools

Client-side tools — functions you define and execute yourself — add no separate charge. They do add tokens, which is a different article: tools cost tokens too.

Checking before you build

server-tool-cost.txt
My feature will use <web search / web fetch / code execution> roughly
<n> times per user action, with about <n> user actions a day.

Work out the monthly cost split into: token cost, and the separate
per-use charges for that tool. Show the arithmetic.

Then tell me whether a cheaper design exists — caching results,
fetching instead of searching, or doing fewer calls per action.

Search is the one to watch. It's the most useful of these and the only one with a meaningful per-call price attached.