Budget your weekend agent's first cost: count successful calls

Your agent can search the web or pull a page, and the first real cost is the call that comes back with something. You do not need a finance dashboard; you need a number you can check after a weekend build. The rule is simple enough to fit on a sticky note.

Most of the agent can run on a key that already covers model calls. The web is the part that reaches outside the machine, and most side projects stay side projects, which is fine. A small, exact budget for agent tooling keeps the project from becoming a surprise. Keep the two separate, and the weekend project stays easy to reason about.

DeepSeek made its Search and Fetch APIs available to all users on October 17, 2026, alongside the DeepSeek-V4.1-Flash-1017 release. DeepSeek began charging for server-side Search and Fetch calls made through DeepSeek Harness, Claude Code, or Codex at 00:00 on November 1, 2026 in UTC+8.

A weekend agent can make a lot of attempts before it finishes one task. It may search, fail, retry, fetch a page, and then fetch another. The log should show which attempts came back with something and which did not. The log is the invoice you can read without waiting for a bill.

Budget the calls you can actually see

Count only successful Search and Fetch calls, multiply by $0.75 per 1,000, ignore failures, empty results, and timeouts, and set a monthly cap.

• Log the calls your agent makes. Done means a sheet with date, endpoint, status, result size, and cost.
• Mark a call billable with a yes column when it returns HTTP 200 and produces a result, such as a non-empty Search results array or Fetch page content.
• Keep the failed calls out of the total: DeepSeek does not charge for unsuccessful Search or Fetch calls, searches that return no results, or timed-out calls.
• Multiply the billable rows by the per-1,000 price. The total is the number you compare against the cap.
• Set a cap before the month starts. The cap is a number you can hit without checking the invoice.

The common mistake is to count every request, even the ones that do not return anything. The budget should follow the result, not the attempt. Check the yes column before you multiply anything.

Make the log cheap to keep

Keep the log in the same folder as the agent. A plain text file works if a spreadsheet feels like too much. The columns do not need to be fancy; they need to survive a tired review. If you cannot tell which call produced the result, you cannot defend the cost.

Review the log after a weekend build, not after every call. You are looking for patterns: repeated searches for the same question, fetches that return nothing useful, and loops that keep asking for more context. Fix those patterns before the next session.

Name the file after the project, not after the date. If you have several agents, keep one log per agent so the cap stays attached to the thing that makes calls.

Keep the loop from becoming a bill

For DeepSeek Search, changing the count value changes the number of returned results without changing the price of the request. Treat the count value as a way to tune result size, not price. DeepSeek's account-level throughput limits are 100 requests per second for Search and 25 requests per second for Fetch, and exceeding them returns HTTP 429. Use those errors as a sign to slow the agent down and add a pause between retries.

Cap the month before it surprises you

Use the total as the likely monthly bill, add it as a budget line, and start with a number you can defend before the next build. Put the cap in the same sheet as the log, at the top next to the project name, so the number is there before you start the next session. If the agent is running unattended, the cap should be the thing that stops it, not a memory you hope you keep.

Set the cap below the number that would make you stop building. When the sheet hits the cap, pause the agent, review the rows, and decide whether the next search is worth the call.