DeepSeek’s 13 November note opens the app’s tools on deepseek-flash, and one search call can be three charges
The changelog dates the release and does not print an hour. Search and fetch stay at $0.75 per thousand on the pricing page. The web-tools guide bills every query inside a call that returns results. A shortened sample on the chatbot page prints one displayed search and two billed searches.
Feature4 minIn-depth
Priya Shah4 min read
- DeepSeek
- API
- Agents
- Pricing
DeepSeek on Friday opened the conversation tools from its own app on the developer API, on the model it calls deepseek-flash. The changelog entry is dated 13 November and does not print an hour. It says web search, file input, URL reading, voice and safety checks, which the app already combines with the model, can now be switched on with parameters. The service runs the tool calls and assembles the context. The news post names the same model as V4.1-Flash and lists the same work across three endpoints: Messages, Responses and Chat Completions. The pricing page’s version string for deepseek-flash is DeepSeek-V4.1-Flash-1017. That page also prints a one-million-token context, a maximum output of 384K, and a concurrency limit of 2,500.
Thinking on or off decides whether a pasted link is read
The split inside search is a switch, not a second product. With thinking disabled and the deepseek_web_search tool declared, the model searches the web or images and answers with citations. With thinking enabled and the same tool declared, it searches and reads pages across several reasoning steps. URL reading is a separate tool, deepseek_read_url. The web-tools guide says only the last user message is checked for links, repeated URLs count once, and at most the first 10 different URLs are considered. The model decides which of those to read. If thinking is enabled and deepseek_web_search is also declared, the guide says the round uses deep thinking search and does not read links separately. The chatbot page’s own example turns thinking on and declares both tools in the same request. File input, on the changelog, accepts PDFs, Office files, images and more, parsed as plain text or as a layout and then added to the conversation.
Safety is a parameter the news post says a public app must turn on, and a duty it says the developer still has. Set safety_check to high and the service reviews input and output and can return a refusal. The news post says public-facing apps must enable the checks, and that developers remain responsible for content moderation, applicable model registration and labeling of AI-generated content. The chatbot guide repeats that list as obligations a public service has to fulfill on its own. A page that fails the review comes back as url_not_allowed. The web-tools guide says the text of a fetched page is not in the result block. It is carried encrypted in a signature. The url, title and publication fields are for display, and changing them before the block is sent back has no effect.
Seventy-five cents a thousand, counted by the query
The dollar figures are on the pricing page, which the Friday post points to rather than reprinting. Search, text or image, is $0.75 per 1,000 searches. Web fetch is $0.75 per 1,000 fetches. File parsing and the OCR API are $0.225 per 1,000 pages. Text to speech is $0.45 per 10,000 characters at list price, and $0.225 in a column the page labels a limited-time price at 50 percent off. Speech to text is $0.116 per hour of audio, and $0.058 in that same column. The table does not print an end date for the discount. Token prices for deepseek-flash, per million, are printed as $0.003 and $0.006 for a cache hit on input, off peak and at peak, $0.15 and $0.3 for a cache miss, and $0.6 and $1.2 for output. Off peak is half of peak. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday, excluding Chinese public holidays. Weekends and those holidays, the page says, are off peak for the full day.
The web-tools guide is where a single search stops being a single charge. A search call that returns results is billed for each of its queries, so three queries count as three. A call that fails or comes back empty is not billed. An image search bills one search per image and one more for each region the model selects. A fetch or a URL read is billed per page fetched successfully, including a page whose result is then replaced by a url_not_allowed or unavailable error under the safety check. Links that cannot be fetched are not billed. Opening a page whose content already arrived with the search results is free, and so is finding text inside a page. The usage object distinguishes the counts a client might show from the counts that are billed. web_search_requests and web_fetch_requests include errors and are not the billed counts. tool_search, tool_fetch and tool_image_search are. The shortened stream on the chatbot page prints both. At the end of that example, web_search_requests is 1 and tool_search is 2. The page does not explain the gap. It does say the first field is for display.
Turning the tools on also drops the sampling knobs. When a request declares web search or URL reading, sets safety to high, or asks for a conversation title, the chatbot guide says the service decides temperature, top_p and a list of other fields, and the values in the request are ignored. On the Messages API that list includes thinking.budget_tokens. max_tokens, called max_output_tokens on the Responses API, still applies, and it limits each inference step in a round, not the round’s total output. The open-source demo, linked from the Friday post, stores the API key, the conversations and uploaded images only in the browser. An import of a chat history exported from the DeepSeek app replaces the history already there. The page says the demo is for demonstration only.



