

The deepseek api cost did not move. The search bill did.
On 17 October 2026 DeepSeek opened a search tool and shipped a faster Flash build, and the deepseek api cost for tokens stayed put. The new bill is per successful call, not per token.
By Katie Delaney / 2026-10-23 / 8 min read

What shipped, and what the deepseek api cost did not do
The deepseek api cost story is a split, not a sale. On 17 October 2026 the lab put DeepSeek-V4.1-Flash-1017 on the website, the app and the API, and it opened Search and Fetch to everyone. The release note says benchmark scores rose, tasks finished faster, and outputs got shorter. It calls that a Pareto improvement: better work at a lower cost of running the task. It does not say the token price changed.
The change log for the same date is plainer. The calling method and the token price stay as they were. You still ask for deepseek-flash. Older names, deepseek-v4-flash and deepseek-v4-flash-vision-exp, are still accepted, and those requests are served by the new build and billed at the Flash price. The first-call page repeats that mapping. A buyer who only reads a price-tracker headline will miss the split.
Treat the score table as the lab's own voice, not an independent exam. The change log lists HLE without tools at 42.6, with a starred 44.0 beside it, Terminal-Bench 4.0 at 42.3, DeepSWE v1.1 at 77.0, ProgramBench at 29.1, NL2Repo-Bench at 68.2, Automation-Bench v1.0.6 at 60.5, Agents' Last Exam at 35.4, Apex Bench v1.1 at 72.1, Chartography with tools at 83.0, BabyVision with tools at 92.5, and ZeroBench-main with tools at 51.0. A chart on the release page was not readable as text here, so those image figures stay out of the prose.
The deepseek search api is a call bill, not a token bill
The deepseek search api is two endpoints on the same base URL. Search is POST https://api.deepseek.com/v1/tools/search. Fetch is POST https://api.deepseek.com/v1/tools/fetch. The Search and Fetch guide bills each at $0.75 per 1,000 successful requests. A request counts only when it returns HTTP 200 and a result: a non-empty results list for Search, page content for Fetch. Failures, empty searches and timeouts are not billed. The count parameter changes how many rows come back. It does not change the price.
The Search reference caps the query at 1 to 128 characters, and it cannot be only spaces. Count runs from 1 to 50 and defaults to 10. A rerank score runs from 0 to 3, and the page calls 1.5 or above high quality. Authority runs from 0 to 4, with minus 1 meaning unknown, and 3 or above marked as highly authoritative. The Fetch reference caps the URL at 8,192 characters, requires http or https with a host, and refuses a username or password in the URL.
| Event | Billed? |
|---|---|
| HTTP 200 with a result | Yes, one request |
| Empty search | No |
| Failure or timeout | No |
| A higher count value | No extra charge |
- HTTP 200 with a resultYes, one request
- Empty searchNo
- Failure or timeoutNo
- A higher count valueNo extra charge
Operators on the guide are the practical part a retrieval team will actually type: site: and -site:, a leading minus for an excluded word, quotes for a phrase, lang: with comma-separated codes, and after: or before: with a year or a full date. That is a search dialect, not a chatbot trick. Write it into the deepseek api documentation your engineers already keep, or the first week of use will be a thicket of near-miss queries.

A free window, then one clock
Server-side Search and Fetch calls made through DeepSeek Harness, Claude Code or Codex are free for now. The guide says per-request billing starts at 00:00 Beijing time on 1 November 2026. The 17 October release note says the same start as 31 October 2026 at 16:00 UTC. Those are one instant, not two deadlines. Beijing is eight hours ahead of UTC, so midnight there is 16:00 the previous day in UTC. Put one clock in the runbook.
| Item | Value |
|---|---|
| 2026-09-10 | Flash release |
| 2026-10-17 | Search opens |
| 2026-10-31 | Call bill starts |
The September note is the backdrop, not today's news. DeepSeek-V4.1-Flash, 10 September said the new architecture cut API prices and put native multimodal support on the API. This October page does not repeat a token-price cut. If a slide deck still says the October update halved tokens, it is ahead of the documents.
How a deepseek api python call should be read
A deepseek api python example on the thinking-mode page uses the OpenAI client, base URL https://api.deepseek.com, model deepseek-flash, and a thinking flag. The thinking-mode guide returns the chain of thought in reasoning_content, beside content. If the request carries tools, that reasoning must be passed back on later turns or the API returns a 400. Without tools, the sample says the next turn ignores reasoning_content.
The tool-calls guide is explicit that the model does not run the function. Your code does. Chat Completions will not let you insert a tool call the model did not generate into the middle of a thread. The Anthropic-format API and the Responses API will. The Anthropic-format guide sets the base URL to https://api.deepseek.com/anthropic. Names starting claude-opus map to deepseek-v4-pro, billed at the V4 Pro price. Names starting claude-haiku or claude-sonnet map to deepseek-flash. An unsupported model name is mapped to deepseek-flash.
That mapping is a cost decision wearing a compatibility coat. A team that points a Claude-shaped client at this base URL and leaves an Opus-style name in the config will pay the Pro price, not the Flash price. Read the model string in the live request, not the nickname in the slide. The release note also says the best Harness experience needs version 0.2.1-rc.1 or later, and it names Tencent CodeBuddy and WorkBuddy as tuned targets. Weights are named on that page at Hugging Face and ModelScope. This run did not download the files, so the links are the lab's own pointers, not a file audit.
For a buyer choosing an agent stack, the useful question is which meter will dominate the month. A short chat still lives on tokens. A retrieval loop that searches and fetches on every step will feel the $0.75 line once the free window closes, even if the token price never moved. That is the whole of the deepseek api pricing change this week: a new line, not a rewritten card. Ask the deepseek api cost question in two parts, or the vulpine habit of reading the small print will do it for you.

What a buyer should do before 31 October
First, find every place the product calls search or fetch through a harness, Claude Code or Codex. Those calls are the ones that turn from free to billed at 16:00 UTC on 31 October. A direct API call is already on the $0.75 meter, if the guide's rule is the one your account follows. Do not assume the harness discount covers a hand-rolled client.
Second, log HTTP status and whether a result came back. You pay for success with content, not for a timeout. A dashboard that counts every attempt will overstate the deepseek api cost. A dashboard that counts only 200s with a body will match the guide.
Third, keep the token card where it was. Nothing fetched this morning says the Flash token price moved on 17 October. If finance asks for a new rate card, show them the change log line that the calling method and the price are unchanged, and show them the new call meter beside it. The deepseek api cost for tokens is the old card. The deepseek api cost for search is the new line. Two numbers, two jobs. Mixing them is how a quiet month becomes a surprise invoice. Follow that trail before the free window closes.
The commercial path from here is ordinary. A firm that wants the retrieval design, the meter, and the runbook written down can start with the AI consultancy page, or go straight to contact. Pricing for the work sits on the pricing page. Search teams who care how answers cite sources should also read the SEO and GEO note, because a search API in the agent is a cousin of the answer box, not a replacement for the site. Paid acquisition still lives on paid search, and the other pieces from today sit on the news stream.
Frequently asked questions
What is the deepseek api cost for search?
The Search and Fetch guide prices each endpoint at $0.75 per 1,000 successful requests. That deepseek api cost covers a successful call only. Empty results, failures and timeouts are not billed.
Did deepseek api pricing change on 17 October 2026?
The change log says the token calling method and the token price were unchanged. What opened was a separate per-call meter for Search and Fetch, plus a new Flash build under the same model name.
Where is the deepseek api documentation for Search?
The guide is at api-docs.deepseek.com/guides/search. The request schemas are on the Search and Fetch reference pages. The 17 October release note restates the $0.75 price and the billing start.
Which model name should a deepseek api python client send?
The first-call page says to send deepseek-flash. That name is served by DeepSeek-V4.1-Flash-1017. Older Flash names are still accepted and billed at the Flash price.
When does the deepseek search api stop being free in a harness?
The release note says billing for harness, Claude Code and Codex tool calls starts at 16:00 UTC on 31 October 2026. The guide's Beijing midnight on 1 November is that same moment.
Bring the meter into the den
If an agent stack needs a written split between token price and call price, that is a consultancy job, not a slogan.
Want folkfox in your Google results and AI answers? Set folkfox as a preferred source.

