DeepSeek, a Hangzhou-based AI lab, opened the search, file-reading, voice and safety functions behind its consumer app to developers through its API on 13 November, Pandaily reported on 14 November 2026. The update to the company’s API change log lets developers put those functions into their own products, using the model through existing endpoints or calling some services separately.
Key points
- The deepseek-flash model can use search, file input, URL reading and safety checks across three existing API endpoints.
- Separate APIs cover Search, Fetch, OCR, speech recognition and speech synthesis.
- Search and file-processing charges are added to token charges; public-facing services must handle their own compliance duties.
- DeepSeek has released an open-source chat-demo that connects to the API and can import chat history from its app.
Deepseek-flash connects search to three endpoints
The model serving these features is deepseek-flash, currently backed by the DeepSeek-V4.1-Flash-1017 checkpoint. Developers can turn functions on with parameters in the Messages, Responses and Chat Completions APIs. DeepSeek’s servers then make the tool calls and put the retrieved material into the context from which the model answers. For a developer already using one of those endpoints, search or file reading can therefore be enabled within the model request rather than assembled solely through separate calls.
With thinking disabled and its web-search tool enabled, deepseek-flash can search the web or images and answer with citations. Turning thinking on lets the same tool search and read links over several reasoning steps. A separate URL-reading tool fetches pages linked in a message. File input extracts text from PDFs, Office documents and images, or parses their layout before adding the contents to the conversation. These are distinct ways of supplying material to an answer: one seeks it online, while the others work from a link or file already supplied.
The expansion follows DeepSeek’s opening of standalone Search and Fetch APIs on 17 October, when it also updated the V4.1-Flash-1017 checkpoint, as AI Affairs reported at the time. The November release brings those search capabilities into a broader set of app functions available to developers.
OCR and voice have separate APIs
DeepSeek also offers standalone Search and Fetch APIs, alongside optical character recognition, or OCR, which turns images of printed text into machine-readable words. Automatic speech recognition, or ASR, turns speech into text, while text-to-speech, or TTS, makes spoken audio from written output. Search finds material on the web and Fetch retrieves the contents of a particular page, Crypto Briefing reported. The distinction matters when an application needs one operation on its own rather than a complete model-generated reply.
Checking a scanned receipt could involve asking about its contents instead of typing out its printed words, if OCR were connected to a conversation that used the recognised text. A spoken question could likewise be transcribed before it reached the model, and a written answer could be read aloud. Each part would depend on how those separate services were connected.
DeepSeek released an open-source static web client called chat-demo to show the functions working together. It connects from the browser to the API and supports search, link reading, attachments, voice input and the thinking mode. It can also import a user’s DeepSeek chat history. For developers, the client provides an example of the interface around the API calls, as well as the calls themselves.
Search and Fetch carry call charges
Tool use is charged in addition to model tokens. Pandaily reported prices from DeepSeek’s pricing page of $0.75 per 1,000 calls each for search and web fetch, whether called directly or used by the model. File parsing and OCR cost $0.225 per 1,000 pages. TTS is listed at $0.45 per 10,000 characters and ASR at $0.116 per hour of audio, with both voice services currently half price. Those units put different kinds of work on different bills: calls for retrieval, pages for files, characters for generated speech and audio time for recognition.
At peak hours, deepseek-flash still costs $0.30 for a million cache-miss input tokens and $1.20 for a million output tokens, while off-peak prices are half the peak rates. Startup Fortune reported that DeepSeek’s guide gave no prices and that it could not verify a rate on the company’s own page.
DeepSeek’s safety check leaves compliance duties
The model-safety parameter examines inputs and outputs, issuing an indication to decline when it detects a potential hazard. DeepSeek tells developers serving the public to enable the check and to meet their own obligations for content review, model registration and AI labelling. The check can be part of a model request, but the obligations attach to the service built around it.
Startup Fortune reported that enabling search, page reading, stricter safety or automatic titles makes the model disregard settings governing output randomness and probability-based token selection. It also reported that the output-token limit applies to each inference step in a round, rather than to the round as a whole. In chat-demo, importing history from the DeepSeek app replaces the history already in the demo, which stores the API key, conversations and uploaded images in the browser.
