DeepSeek API Adds Web Search, File Parsing, OCR, and Voice

DeepSeek opened web search, deep thinking search, URL reading, file input, and a safety check to developers on deepseek-flash across its Messages, Responses, and Chat Completions endpoints in its November 13, 2026 Change Log, alongside standalone Search and Fetch, OCR, speech-to-text, and text-to-speech APIs billed on top of tokens.

What Did DeepSeek Announce on November 13, 2026?

DeepSeek published a November 13, 2026 entry in its official Change Log titled “Releasing the full suite of application APIs,” bringing the capabilities behind the DeepSeek App and web client to developers. DeepSeek stated that web search, file input, URL reading, voice interaction, and safety already run alongside model inference inside its app, adding: “Today we open these capabilities to developers,” with tool execution and context assembly managed by the service. DeepSeek also shared the release on its official X account on November 13 Beijing time.

The rollout covers five integrated capabilities for deepseek-flash chats and four standalone utility APIs: Search and Fetch, OCR, automatic speech recognition (ASR), and text to speech (TTS). This follows an October 17, 2026 Change Log update in which DeepSeek introduced its in-house search API (DeepSeek Search), which it says indexes hundreds of billions of web pages, alongside the DeepSeek-V4.1-Flash-1017 release under the deepseek-flash model name.

To demonstrate the endpoints, DeepSeek open-sourced a browser-based chat client (live demo and GitHub repository) that can import exported DeepSeek chat histories. DeepSeek notes that API keys, conversations, and uploaded images stay inside the browser and that the client is for demonstration only.

How Do the Five In-Chat Capabilities Work on deepseek-flash?

All five in-chat capabilities run on deepseek-flash (currently listed as DeepSeek-V4.1-Flash-1017 on DeepSeek’s pricing page) across three API endpoints:

• Messages (Anthropic format): POST https://api.deepseek.com/anthropic/v1/messages
• Responses: POST https://api.deepseek.com/responses
• Chat Completions: POST https://api.deepseek.com/chat/completions

According to the Build a Chatbot guide and Web Tools guide, each capability maps to specific request settings:

• Web search: With thinking disabled and the deepseek_web_search tool enabled, deepseek-flash runs web or image searches as needed and answers with citations. Image search triggers automatically when the user message contains images.
• Deep thinking search: With thinking enabled alongside deepseek_web_search, the model searches the web and reads links autonomously across multiple reasoning steps, making up to roughly 20 tool calls before answering.
• URL reading: Enabling deepseek_read_url lets the model detect links in the user prompt, fetch those pages, and include them in context when answering.
• File input: Uploading PDFs, Office documents, images, and other files through the Files API parses them into context using either text_recognition (plain text only, dropping charts and photos) or rich_layout_recognition (text plus cropped figures placed at their original positions).
• Model safety: Adding safety_check screens inputs and outputs and returns a refusal marker when risk is detected.

The chatbot guide also describes optional user metadata markers (passing current date, user location, and language) and automatic conversation title generation via title. Client applications must display the returned content blocks and send them back unchanged on the next turn. Because DeepSeek handles search calls and context assembly server-side, this workflow contrasts with traditional client-side tool calling, where your application executes every tool call locally.

How Do You Enable DeepSeek API Web Search in a Request?

Developers enable DeepSeek web search, URL reading, high-level safety screening, and automatic conversation titles by adding those parameters directly to a deepseek-flash JSON request body:

{
  "model": "deepseek-flash",
  "tools": [
    { "type": "deepseek_web_search" },
    { "type": "deepseek_read_url" }
  ],
  "safety_check": { "level": "high" },
  "title": { "enabled": true },
  "messages": [
    {
      "role": "user",
      "content": "Summarize the latest pricing updates on https://api-docs.deepseek.com/quick_start/pricing"
    }
  ]
}

When a request declares the web tools, sets safety_check to high, or enables title, DeepSeek’s service overrides several sampling parameters and ignores client values such as temperature and top_p. The max_tokens parameter remains active, though it caps each individual inference step rather than the entire multi-step search round.

Standalone DeepSeek Search, Fetch, OCR, ASR, and TTS APIs

Alongside the integrated chat features, DeepSeek opened four standalone APIs so developers can invoke search, page extraction, OCR, and voice processing outside a chat session.

Search and Fetch APIs

Described in the Search & Fetch guide, POST https://api.deepseek.com/v1/tools/search queries DeepSeek’s web index directly, while POST https://api.deepseek.com/v1/tools/fetch retrieves a web page as Markdown. Unlike the in-chat deepseek_web_search tool, the standalone Search API does not support image search yet.

DeepSeek OCR API

Exposed at POST https://api.deepseek.com/v1/tools/ocr, the OCR API supports PDF, Word, PowerPoint, and common image formats, returning extracted text page by page. Files sent to the OCR endpoint are neither stored by DeepSeek nor passed into a chat model.

DeepSeek ASR (Speech to Text)

Available at POST https://api.deepseek.com/v1/tools/asr, the ASR API transcribes whole 16-bit PCM WAV or Ogg Opus audio files in a single request, with no streaming recognition. It automatically detects 21 languages, including Chinese, Cantonese, English, Japanese, Korean, German, French, Spanish, Portuguese, Italian, Russian, Arabic, and Hindi.

DeepSeek TTS (Text to Speech)

Available via POST https://api.deepseek.com/v1/tools/tts for full-text synthesis or WebSocket streaming, the TTS API offers two voices: stella (female, default) and tide (male). It auto-detects 10 languages (Mandarin, Cantonese, English, Japanese, Korean, Malay, Arabic, Indonesian, Thai, and Vietnamese) and outputs wav by default or pcm.

What Does DeepSeek API Web Search, OCR, and Voice Cost?

As of November 15, 2026, DeepSeek bills all tool and standalone API calls in addition to deepseek-flash model token fees, according to DeepSeek’s official Models & Pricing page.

Tool and standalone API prices

Item	Price	Unit	Notes
Search (text and image)	$0.75	Per 1,000 searches	Same price via web tools or Search API
Web fetch	$0.75	Per 1,000 fetches	Same via URL reading or Fetch API
File parsing and OCR API	$0.225	Per 1,000 pages	Charged per parsed page
Text to Speech (TTS)	$0.45 list / $0.225 limited-time	Per 10,000 characters	50% off limited-time price
Speech to Text (ASR)	$0.116 list / $0.058 limited-time	Per hour of audio	50% off limited-time price

deepseek-flash token prices (per 1M tokens)

Token type	Off-peak	Peak
Input (cache hit)	$0.003	$0.006
Input (cache miss)	$0.15	$0.30
Output	$0.60	$1.20

deepseek-flash supports a 1M token context window and up to 384K output tokens. Off-peak token prices are half of peak rates. Peak hours run Monday through Friday from 01:00 to 04:00 UTC and 06:00 to 10:00 UTC, excluding Chinese public holidays; all other hours and weekends use off-peak rates. DeepSeek has not announced an end date for the 50% TTS and ASR discount and states that prices may vary and are subject to adjustment. For broader token-rate context across providers, see our AI API pricing comparison.

Which Billing Rules Surprise Developers Most?

DeepSeek’s billing model ties charges to individual sub-queries, successful fetches, uploaded pages, and safety review tokens rather than raw endpoint calls:

• Search bills per query rather than per tool call: A single deepseek_web_search invocation containing three search queries counts as three billed searches. For image search, each image counts as one search, plus one more for every region the model selects.
• Failed and empty results are free: Failed searches, empty results, timeouts, and links that fail to fetch incur no fee. Standalone Search and Fetch requests only bill on HTTP 200 responses with non-empty results, regardless of result count or page tokens.
• Deep thinking search reopens cost nothing: During multi-step reasoning, a page is billed only when fetched from the web. Reopening a fetched page, moving through it, or finding text inside it carries no extra fetch charge.
• Safety checks bill as output tokens: Tokens used for safety review are reported as safety_tokens and charged at output token rates.
• File parsing bills once at upload: Document parsing charges per page once upon successful upload; subsequent chat turns referencing the file pay standard input token rates.
• Voice billing: ASR bills audio duration to the second.

Key Limits, Safety Checks, and Compliance Obligations

DeepSeek enforces strict payload caps, tool restrictions, and compliance caveats across the new endpoints.

Technical and Concurrency Limits

• Web tools: allowed_domains, blocked_domains, and user_location are not supported on deepseek_web_search. Other server tools cannot be combined with deepseek_web_search, deepseek_read_url, safety_check, or title yet. Deep thinking search stops after about 20 tool calls.
• File parsing and OCR: Both cap files at 100 MiB and 1,000 pages. File parsing enforces a 120-second upload limit, a 120-second parsing limit, and 15 concurrent uploads per account; original files and parsed outputs cannot be downloaded. Standalone OCR allows 180 seconds total.
• ASR and TTS: ASR limits files to 8 MiB and 120 seconds, with up to 50 concurrent requests per account. TTS caps requests at 10,000 billed characters and does not support speed control.

Model Safety Checks and Public Service Compliance

By default, the API runs low level safety checks. Passing safety_check set to {"level": "high"} invokes the Model Content Safety review across both inputs (including tool results) and outputs. Violations return 400 Content Exists Risk or end the response with a refusal (stop_reason "refusal" in Messages; finish_reason "content_filter" in Chat Completions). Because already-streamed tokens cannot be withdrawn, clients must replace streamed text with their own refusal message when a refusal arrives.

DeepSeek instructs operators providing public-facing services to enable safety_check and fulfill obligations including content safety review, large model registration, and AI labeling. DeepSeek notes that safety_check aids compliance without replacing a service provider’s own content review duties.

When Should You Use In-Chat Web Tools vs. Standalone Search and Fetch?

Choosing between DeepSeek’s in-chat web tools and its standalone Search and Fetch APIs depends on whether you want DeepSeek or your own code controlling retrieval. Use deepseek_web_search and deepseek_read_url inside deepseek-flash when building conversational assistants that benefit from server-managed search orchestration, image search, multi-step deep thinking, and automatic citations.

Use the standalone /v1/tools/search and /v1/tools/fetch endpoints when building custom retrieval-augmented generation (RAG) pipelines that require your own domain filtering, chunking, or downstream model selection over Markdown page extracts. Standalone OCR, ASR, and TTS similarly fit modular pipelines where files or audio do not belong in chat state.

FAQ

Which DeepSeek models support web search and URL reading?

DeepSeek’s documentation specifies deepseek-flash (currently listed as DeepSeek-V4.1-Flash-1017 on the pricing page) across the Messages, Responses, and Chat Completions endpoints. The guides do not list support for other model IDs, so developers needing server-handled web search, deep thinking search, URL reading, file input, and safety checks should target deepseek-flash.

Can you restrict DeepSeek web search to specific domains?

No. DeepSeek’s Web Tools guide confirms that allowed_domains, blocked_domains, and user_location are not supported on deepseek_web_search. If you need strict domain allowlists or blocklists, call the standalone /v1/tools/search API, filter the returned URLs in your own backend, and retrieve approved pages using /v1/tools/fetch.

Are failed searches or unreachable URLs billed on the DeepSeek API?

No. DeepSeek bills web search only when a query returns results and bills URL reading only when a page fetches successfully. On the standalone Search and Fetch APIs, timeouts, failed requests, and empty results incur no charge; only HTTP 200 responses with non-empty payloads are billed at $0.75 per 1,000 requests.

How long will DeepSeek’s 50% discount on TTS and ASR last?

DeepSeek’s Models & Pricing page lists a limited-time 50% discount lowering TTS to $0.225 per 10,000 characters and ASR to $0.058 per hour of audio, without stating an end date. DeepSeek also states that prices may vary and reserves the right to adjust published API rates.

Does enabling safety_check satisfy public AI compliance rules by itself?

No. DeepSeek states that operators offering a service to the public must enable safety_check and complete obligations covering content safety review, large model registration, and AI labeling. The documentation specifies that the parameter helps with compliance without replacing a service provider’s own content review responsibilities.

Sources

• DeepSeek API Docs, Change Log
• DeepSeek API Docs, Models & Pricing
• DeepSeek API Docs, Build a Chatbot
• DeepSeek API Docs, Web Tools
• DeepSeek API Docs, Search & Fetch
• DeepSeek API Docs, Model Content Safety
• DeepSeek on X