News · 2026-11-15
DeepSeek documents search, file and voice APIs for developers
DeepSeek’s November 13 API update documents developer access to hosted search, URL reading, file processing, safety checks and voice tools associated with its app. The expansion lets applications combine model conversations with services run by DeepSeek, while adding separate tool charges and data-handling decisions.
Key facts
- The release note is dated November 13, 2026.
- Search and fetch are priced at $0.75 per 1,000 calls, separately from applicable model-token charges.
- File expiration is optional; the Files API says omitted expiry means a file is kept permanently.
- The primary source is DeepSeek’s release announcement.
A chatbot that can answer from current web pages needs more than a language model. It must choose a search, retrieve a page, convert it into usable text and place that text in the conversation. DeepSeek’s web tools guide describes an integrated path in which deepseek-flash can invoke search or URL reading and DeepSeek runs the tools on its service. The documentation also describes direct Search and Fetch endpoints for applications that want to make those calls themselves.
Think of the integrated approach as asking a research assistant to go to the library and bring relevant passages back. The direct endpoints are more like giving the application its own library card: the developer decides what to fetch and how to use it. Both approaches require attention to the reliability and trustworthiness of outside material. A retrieved page can be mistaken or contain instructions intended to manipulate the assistant; the prompt injection lesson explains that distinction.
DeepSeek’s file guide describes uploading files and referencing their identifiers in a conversation. Parsing can return text or a richer layout with detected figures, though the documentation does not guarantee every figure will be detected. Its separate speech recognition and speech synthesis guides describe full-audio transcription and streamed audio output respectively. Speech recognition is not documented as streaming input.
The practical cost is a combination of model use and tool use. DeepSeek’s pricing page lists search and fetch at $0.75 per 1,000 calls each; other processing is billed separately. Prices and promotional voice rates may change. A developer should estimate the number of searches and retrieved pages per user task, rather than treating the model’s token price as the complete bill.
The Files API reference says expiry can be set from one hour to 30 days and that the file is kept permanently if the parameter is omitted. That is an important default to review before uploading confidential material. It does not by itself establish all logging, training-use, storage-location or deletion terms for prompts and results. DeepSeek also describes its high safety setting as “basic risk control,” and says public-facing developers retain their own responsibilities.
The update does not mean every endpoint is new. The company’s changelog already announced direct Search and Fetch in October. The dossier found documentation and published prices but no authenticated production test or independent adoption evidence. The news is the documented API surface and its orchestration options, not proof of worldwide availability or of a new underlying model for every task.
File processing is another division of labor. DeepSeek says a supported file is parsed when uploaded and the returned file identifier can be referenced in a conversation. Its separate optical text recognition endpoint handles a file directly rather than storing it in the Files API or passing it to a chat model, according to the documentation. These two paths can have different data-handling implications even when both begin with the same document.
Developers should also distinguish the existence of a documented endpoint from demonstrated fitness for a particular use. A transcription result can contain mistakes; a retrieved page can be stale or adversarial; a parsed document can omit a chart; a safety check can miss content or block a legitimate request. The documentation describes available controls and limits, but the dossier does not contain an independent evaluation of quality across those tasks.
The launch matters because packaging services around a model changes the work an application developer must do. A hosted search tool can reduce the need to operate a separate retrieval pipeline, while direct endpoints offer more control over when calls happen and what results are included. The cost and control choices are connected: an application making several searches and fetches per answer pays for each applicable call, and any externally retrieved content still needs to be treated as untrusted input.
Key questions
Does DeepSeek’s API run web searches itself?
Are uploaded files deleted automatically?
Was every search endpoint introduced on November 13?
Cite this
APA
Ground Truth. (2026, November 15). DeepSeek documents search, file and voice APIs for developers. Ground Truth. https://groundtruth.day/news/deepseek-expands-app-api-search-files-voice.html
BibTeX
@misc{groundtruth:deepseek-expands-app-api-search-files-voice,
title = {DeepSeek documents search, file and voice APIs for developers},
author = {{Ground Truth}},
year = {2026},
month = {nov},
url = {https://groundtruth.day/news/deepseek-expands-app-api-search-files-voice.html}
}
Comments are replies to this story on Bluesky — reply with any Bluesky account to join in.