DeepSeek put its app's search in the API at $0.75 per 1,000 calls. The model decides how many calls you pay for.
On November 13 DeepSeek moved search, file parsing and link reading from its chat app into the API. The per-call price is low. The catch is who controls the call count.
Twenty. That is the number I’d keep in mind from DeepSeek’s November 13 release, because it decides what the 75-cent price tag turns into on a real invoice.
DeepSeek’s announcement says the capabilities behind its consumer app are now in the API: cited web search, a deeper search mode, file input for PDFs and Office documents, automatic reading of links, safety checks, and standalone OCR, speech recognition and text-to-speech endpoints. The built-in features run on deepseek-flash, and a developer switches them on by declaring tools called deepseek_web_search and deepseek_read_url. DeepSeek’s web tools guide says the model decides whether to go online, and the searching happens on DeepSeek’s servers, not in your code.
Where the 20 comes from
With thinking mode off, the model searches when a question needs it and answers with citations. With thinking mode on, the same tool becomes what DeepSeek calls deep thinking search. WebProNews, summarizing the guide, reports that the model can make up to about 20 tool calls before it answers, moving between searching, opening pages and finding text inside them. I couldn’t confirm that exact limit on the guide pages I could read, so treat 20 as a ceiling reported by WebProNews, not a number DeepSeek printed in the announcement.
The price per call is on DeepSeek’s pricing page: $0.75 per 1,000 successful search requests and the same for page fetches. The Search and Fetch guide adds that failed requests, empty results and timeouts aren’t billed, and that the model-driven web tools are billed at those same prices. Tokens are charged on top, as before.
What a research run costs
My arithmetic, using only the list prices: 20 calls at $0.75 per 1,000 is 1.5 cents. Anthropic’s web search documentation lists $10 per 1,000 searches plus token costs, so 20 searches there come to 20 cents before tokens. That’s a gap of about 13 times on the tool fee alone.
| DeepSeek API | Claude API | |
|---|---|---|
| Web search, per 1,000 | $0.75 | $10, plus tokens |
| Page fetch, per 1,000 | $0.75 | no per-use fee, according to WebProNews |
| One 20-call run (searches only) | about $0.015 | about $0.20 |
| 10,000 such runs | about $150 | about $2,000 |
The last two rows assume every call is a search. Real runs mix searches and page opens, and on DeepSeek both are billed, so the true figure sits somewhere around that line. Check the vendors’ pricing pages before you budget, since DeepSeek says it reserves the right to change prices.
The part the headline skips
Search itself isn’t new here. DeepSeek opened Search and Fetch as standalone endpoints with its October model release, which I covered in the V4.1-Flash-1017 post. What changed on November 13 is that the model can call them for you, inside one request, without your code seeing each step.
That moves the bill from something you count to something the model decides. If you call a search endpoint yourself, you know you made one call. If you enable deep thinking search and ask a hard question, you find out afterward how many the model chose to make.
My read: at $0.75 per 1,000 that’s a small risk for a hobby project and a real line item for an app serving thousands of users a day, so I’d cap usage in my own code and log the tool counts from the first week. Whether the real average run lands near five calls or near twenty is something only usage data will show, and nobody has published it yet.
DeepSeek also attaches conditions. Public-facing apps must enable its safety checks, and developers “remain responsible for content moderation, applicable model registration, and AI-generated content labeling,” in the company’s words. The speech prices are half-price for a limited time, with no end date given.
Sources
Note on sources: I read DeepSeek’s announcement and pricing page, the billing and rate-limit sections of the Search and Fetch guide, the opening sections of the web tools guide (the rest was cut off in the copy I could fetch), and the pricing line on Anthropic’s web search page. WebProNews is a secondary outlet; its “up to about 20 tool calls” and Claude fetch-fee statements are its reading of those docs, and I did not verify them on the vendor pages. OpenAI’s own search pricing wasn’t checked, so this post leaves it out.