DeepSeek has published an “all-in-one app API” in its API changelog, opening to developers the capabilities its own app uses in conversations: web search, deep-thinking search, file input and link reading. Developers declare the relevant tool or parameter in the existing interface, and the server handles tool calls and context assembly, so there is no need to build your own retrieval, scraping and prompt-stitching pipeline.
Only deepseek-flash can use these features. The pricing page lists the version as DeepSeek-V4.1-Flash-1017, and all three interfaces, Messages, Responses and Chat Completions, support it. The new capabilities do not list deepseek-v4-pro, and the pricing page still shows Pro without image understanding.
How to switch on the four chat capabilities
- Web search: in non-thinking mode, declare deepseek_web_search in tools. The model starts web or image searches on its own and the answer carries citations.
- Deep-thinking search: the same tool combined with thinking mode. The model searches and reads by itself during multi-step reasoning, and answers after at most about 20 tool calls.
- Link reading: declare deepseek_read_url. The model detects links in the user’s message and looks at up to the first 10 distinct URLs.
- File input: upload through the Files API first, then reference the file_id. PDF, Word, PowerPoint and common image formats are supported, with a limit of 100 MiB and 1,000 pages per file.
The documentation lists several restrictions. tool_choice cannot force either of the two web tools, so the model decides whether to search. Filter fields such as allowed_domains and blocked_domains return a 400 error outright, and a reply that used web tools cannot be continued. There is also a safety_check parameter: at the high level, a detected risk returns a refusal flag, which is aimed at applications that need content moderation.
Five features split out as standalone endpoints
Search, Fetch, OCR, ASR and TTS can also be called on their own, outside a conversation, each with its own pricing:
| Capability | Price | Notes |
|---|---|---|
| Search | 5 yuan per 1,000 calls | Up to 50 results per call, QPS 100 |
| Fetch | 5 yuan per 1,000 calls | Only successful requests are billed, QPS 25 |
| OCR / file parsing | 1.5 yuan per 1,000 pages | Same price as file input |
| ASR | 0.78 yuan per hour of audio | Limited-time half price: 0.39 yuan |
| TTS | 3 yuan per 10,000 characters | Limited-time half price: 1.5 yuan |
Fetch returns a title plus Markdown body text, which means the page-cleaning step is covered too. The pricing page also sets out peak and off-peak rules: on weekdays, 9:00-12:00 and 14:00-18:00 Beijing time count as peak, and off-peak prices are half the peak price. Whether tool calls follow the same rule is not stated separately in the documentation.
A rough calculation: a customer-service bot with web access handles 10,000 conversations a day, and each one triggers on average 2 searches and 1 page read. That is about 30,000 calls times 5 yuan per 1,000, or roughly 150 yuan a day in tool fees, with the model’s own token charges on top. What that money replaces is a whole stretch of engineering: buying a search service, maintaining scraping and body-text cleaning, and stitching the results back into the context. For a team of a few people, the main saving is labor.
An open-source web client comes along
DeepSeek has also released chat-demo on GitHub under the MIT license. It is a purely static chat front end: the browser connects directly to the DeepSeek API with no self-hosted backend. It supports multiple conversations, deep thinking, smart search, link reading, attachment upload, voice input and read-aloud, and the API key and settings are stored locally in the browser. In other words, a developer can run a web client close to the official experience locally with their own key.
The changelog says it “supports importing DeepSeek chat history”, while the repository README describes importing and exporting its own ZIP archives, with import overwriting all local conversations. Whether it can read chat records exported from the official app is described inconsistently in the two places. For now the README is the safer guide, and the point still needs hands-on testing.
Sources: DeepSeek API changelog, DeepSeek pricing page (used to check the search, fetch, file parsing and speech price tiers), CocoLoop, GitHub chat-demo repository README.