DeepSeek Application APIs: Power-User Guide to Web Search, Files, OCR and Voice

Last updated: November 14, 2026

DeepSeek’s application APIs, released November 13, 2026, let the deepseek-flash model use web search, deep thinking search, file input, URL reading and a safety check inside the existing chat APIs. DeepSeek also opened standalone OCR, speech-to-text and text-to-speech endpoints, billed per use on top of normal token charges.

DeepSeek describes the release as “the full DeepSeek App experience, now via API” in its November 13 announcement. This guide covers what changed, how to turn each feature on, and the limits and prices.

Contents

What is new

According to the DeepSeek API change log, the deepseek-flash model now supports these features in the Messages, Responses and Chat Completions APIs, with the tool calls and context assembly handled by DeepSeek’s servers:

Feature How it is enabled What it does
Web search deepseek_web_search tool, thinking disabled Runs web or image searches as needed and answers with citations
Deep thinking search deepseek_web_search tool, thinking enabled Searches, opens pages and finds text in pages over multiple reasoning steps
File input Upload to the Files API, then reference the file_id Parses PDFs, Office files and images into conversation context
URL reading deepseek_read_url tool Fetches links found in the user message and answers with them in context
Safety check "safety_check": {"level": "high"} Reviews input and output and returns a refusal when a risk is found
Conversation title "title": {"enabled": true} Generates a short title when the answer ends

Search, Fetch, OCR, speech-to-text (ASR) and text-to-speech (TTS) are also available as standalone APIs. DeepSeek published an open-source demo chat client built on these APIs that can import chat history exported from DeepSeek Chat.

On October 17, DeepSeek opened its search engine only as separate Search and Fetch APIs (see our DeepSeek Search API guide). Your code no longer has to run the search, fetch pages and assemble context itself.

How to enable the chat features

The Build a Chatbot guide lists three endpoints, all used with deepseek-flash: /anthropic/v1/messages, /responses and /chat/completions on https://api.deepseek.com.

  1. Create an API key on the DeepSeek platform and set DEEPSEEK_API_KEY.
  2. Set model to deepseek-flash.
  3. Choose thinking on or off. In Chat Completions and Messages use "thinking": {"type": "enabled"}; in Responses use "reasoning": {"effort": "high"}.
  4. Add {"type": "deepseek_web_search"} and, if you want links read, {"type": "deepseek_read_url"} to tools. In the Messages examples, each tool also carries a matching name.
  5. For a public-facing app, add "safety_check": {"level": "high"}.
  6. Stream the response, show the thinking, search, page and answer blocks, and pass the returned blocks back unchanged in the next request.

When any of these features is on, the service sets some sampling parameters itself and ignores your values, including temperature and top_p in all three protocols. max_tokens still works, but it limits each inference step, not the whole round.

Web search and URL reading

The Web Tools guide gives these rules:

  • One search call can hold several queries; a call that returns results is billed once per query.
  • Deep thinking search makes up to about 20 tool calls before it answers.
  • URL reading checks only the last user message, considers at most the first 10 different URLs, and lets the model decide which to read. Long pages are truncated when several links are read in one round.
  • Read pages come back encrypted in a signature field. Passing the block back unchanged restores the page into later turns.
  • You cannot force the tools with tool_choice, set allowed_domains, combine them with other server tools, or continue a reply that used them.

Answers carry citations, and the usage object reports billed counts in tool_search, tool_fetch and tool_image_search.

File input and the OCR API

File input works in two steps, per the File Parsing guide. Upload the file to POST /files with a purpose field sent before the file field, then reference the returned file_id in a chat message.

  • text_recognition gives the model text only and drops charts and photos.
  • rich_layout_recognition also crops figures and places them in the text, which uses more input tokens.
  • Limits: 100 MiB, 1,000 pages, a 120-second upload window and 120 seconds of parsing. A request can reference up to 20 documents totaling 8 MiB of text.
  • Plain text, spreadsheets and ebooks use purpose=user_data and are not billed.

The standalone OCR API at POST https://api.deepseek.com/v1/tools/ocr returns text page by page without storing the file or sending it to a chat model. It accepts the same formats, allows 180 seconds in total and up to 15 recognitions in progress per account.

Speech-to-text and text-to-speech

The speech-to-text API at POST /v1/tools/asr takes one 16-bit PCM WAV or Ogg Opus file of up to 8 MiB and 120 seconds, detects the language automatically from 21 supported languages, and returns the full transcript. Streaming recognition is not supported.

The text-to-speech API has two modes:

Mode Endpoint Notes
Full text POST /v1/tools/tts JSON body; WAV (default) or raw 24 kHz PCM; up to 30 minutes of audio per request
Streaming wss://api.deepseek.com/v1/tools/tts Send text in chunks as the model writes; PCM or Opus frames; sessions up to 30 minutes

There are two voices, stella (female, default) and tide (male), and 10 languages. Each request or session is limited to 10,000 billed characters, and speed control is not supported. Both voice APIs allow 50 requests in progress per account.

Safety checks and compliance

With safety_check set to high, DeepSeek reviews the input and output, per its model content safety guide. A failed input review returns 400 Content Exists Risk; a failed output review ends the response with a refusal, after text may already have streamed, so your client must remove it. Review tokens are billed as output tokens.

DeepSeek states that public-facing services should enable the check and remain responsible for content review, model registration and AI labeling where required.

Pricing

Tool fees are added to normal deepseek-flash token charges, according to Models & Pricing:

Item Unit Price
Search (text, image) Per 1,000 searches $0.75
Web fetch and URL reading Per 1,000 fetches $0.75
File parsing and OCR Per 1,000 pages $0.225
Text-to-speech Per 10,000 characters $0.45 list, $0.225 limited-time
Speech-to-text Per hour of audio $0.116 list, $0.058 limited-time

For token rates, including peak and off-peak pricing, see our DeepSeek API pricing answer. Concurrency rules are covered in DeepSeek API rate limits, and the model itself in our DeepSeek V4.1 Flash guide. For a comparable cloud offering, see the AgentCore Web Search API guide.

Frequently asked questions

What are the DeepSeek application APIs?

They are the features behind the DeepSeek app, opened to developers on November 13, 2026: web search, deep thinking search, file input, URL reading and a safety check inside the chat APIs, plus standalone Search, Fetch, OCR, speech-to-text and text-to-speech endpoints.

How do I enable web search in the DeepSeek API?

Use the deepseek-flash model and declare the deepseek_web_search tool in the tools array. With thinking disabled the model runs ordinary web search with citations; with thinking enabled it runs deep thinking search, which can search, open pages and find text over several steps.

Does DeepSeek have a text-to-speech API?

Yes. POST https://api.deepseek.com/v1/tools/tts converts up to 10,000 billed characters per request into speech, with the voices stella and tide. A WebSocket version streams audio while text is still arriving. The listed price is $0.45 per 10,000 characters, currently $0.225 at a limited-time 50% discount.

How much does the DeepSeek OCR API cost?

DeepSeek lists file parsing and the OCR API at $0.225 per 1,000 pages. OCR is billed for pages recognized successfully; a recognition that fails or times out is not billed, but a disconnected request that still completes is billed.

Which models support the new DeepSeek chat features?

DeepSeek says the built-in web search, file input, URL reading and safety features are supported by deepseek-flash, currently served by DeepSeek-V4.1-Flash-1017, in the Messages, Responses and Chat Completions APIs.

Sources

Related answers

2 thoughts on “DeepSeek Application APIs: Power-User Guide to Web Search, Files, OCR and Voice”

Leave a Comment