DeepSeek priced Search and Fetch at $0.75 per 1,000 successful calls

On October 17, DeepSeek changed the checkpoint behind deepseek-flash and put a posted price on Search and Fetch, two calls that look up a live web page. What AI help costs is where the hour, the month, and the salary sit side by side. This note adds the new tool line, and the one path that is still free.

What changed under the same name

DeepSeek’s API changelog for October 17 says V4.1 Flash is now DeepSeek-V4.1-Flash-1017 on the website, the app, and the API. The calling method and the model price are unchanged. A call to deepseek-flash reaches the new checkpoint.

The current pricing page says the older names still work. deepseek-v4-flash and deepseek-v4-flash-vision-exp are served by DeepSeek-V4.1-Flash-1017 and billed at the flash price. A process can land on the new checkpoint with no edit to the model string.

DeepSeek says this checkpoint runs tasks in less time and writes shorter outputs. On the October 17 changelog, Terminal-Bench 4.0 is 42.3 and DeepSWE v1.1 is 77.0. That entry says Terminal-Bench 4.0 was tested with the Claude Code framework, and DeepSWE with the Mini-SWE-V2 framework. The September 10 entry on the same page lists 31.2 and 74.2 for those two tests. The Hugging Face card sets 42.3 beside 31.2 and 77.0 beside 74.2, and its note names Claude Code and Mini-SWE-V2 for those tests. Those scores are company-reported. The card has more rows. Read them as DeepSeek’s own measurements.

The changelog asks for DeepSeek Harness 0.2.1-rc.1 or later for the best experience with this model. The model card on Hugging Face licenses the repository and the weights under MIT.

What a lookup costs

Search takes a query and returns structured results for web pages. Fetch takes one URL and returns the page as Markdown. DeepSeek’s Search and Fetch guide prices each at $0.75 per 1,000 successful requests.

A request is billable when it returns HTTP 200 and something usable. Search has to come back with a non-empty results list. Fetch has to come back with page content. Empty searches, failed calls, and timeouts are not billed. Setting count anywhere from 1 to 50 changes how many rows you get. It does not change the price of that request.

That rate is for Search and Fetch when your own code calls them. The guide sets apart a second path. Server-side Search and Fetch calls, made while the DeepSeek API runs through DeepSeek Harness, Claude Code, or Codex, are free for now. The guide starts billing for those calls at 00:00 Beijing time on November 1, 2026. DeepSeek’s post on October 17 sets the start at 16:00 UTC on October 31. Beijing time is eight hours ahead of UTC, so the two lines are one moment.

The guide also says the site you fetch can refuse the request. A refusal comes back as a failure, so it is not billed, and you still do not have the page.

Model tokens are a separate charge. The October 17 changelog says that model price did not change.

Price one ordinary day

Take a process that searches, then opens a few of the pages, and that calls Search and Fetch directly. Give it 200 successful Search calls and 50 successful Fetch calls in a day.

Search is 200 / 1,000 × $0.75 = $0.15. Fetch is 50 / 1,000 × $0.75 = $0.0375. The tool line for that day is $0.1875.

Run that same day 30 times and the tool line is $5.625. Tokens for deepseek-flash sit on top of that, at the posted model rates. If those lookups instead run as server-side tools inside DeepSeek Harness, Claude Code, or Codex, the tool line is $0 until 16:00 UTC on October 31. The token line is not part of that free window.

The $5.625 figure is arithmetic on DeepSeek’s posted rate, for a flat month of successful calls. Put your own two counts into the same formula. A log that mixes failures into the total will overstate the bill, because failures are not charged.

Which path the process is on

Sort each lookup before you forecast it:

• Your code calls Search or Fetch itself. The posted price applies to each successful call.
• Your code calls the DeepSeek API through DeepSeek Harness, Claude Code, or Codex, and the lookup is a server-side tool. DeepSeek says that tool call is free until the moment above, then the same $0.75 rate applies.
• The log cannot separate a successful call from a failed one. Count one day by hand before you extend it to a month.

What to do this week

1. Find every process that sends deepseek-flash, deepseek-v4-flash, or deepseek-v4-flash-vision-exp.
2. Mark each lookup as a direct Search or Fetch call, or as a server-side tool inside DeepSeek Harness, Claude Code, or Codex.
3. From one normal day on the direct path, count successful Search calls and successful Fetch calls. Multiply each count by $0.00075.
4. Write 16:00 UTC on October 31 beside any of the three agent paths. That is when DeepSeek says those tool calls start to bill.
5. If you use DeepSeek Harness, run that process on 0.2.1-rc.1 or later and repeat one job whose right answer you already know. Keep DeepSeek’s score table next to that result, and use the job you can check.