DeepSeek opens its app's search, file, voice and safety features to the API
DeepSeek now exposes the features behind its chat app through its API: server-side web search, URL reading, file parsing and safety checks on deepseek-flash, plus standalone Search, Fetch, OCR, ASR and TTS endpoints.
DeepSeek has opened the features behind its consumer chat app to developers. In a news post dated 2026-11-13, the company says its deepseek-flash model (V4.1-Flash) can now run web search, multi-step "deep thinking" search, file input, URL reading and safety checks on the server side, across all three request formats it supports: Messages, Responses and Chat Completions. Search, Fetch, OCR, speech-to-text (ASR) and text-to-speech (TTS) are also sold as standalone APIs.
Alongside the docs, DeepSeek published an open-source chat client, deepseek-ai/chat-demo (MIT licence, created on 2026-11-12, about 1,900 GitHub stars when we checked on 2026-11-15). It can import chat history exported from DeepSeek Chat.
What changed for developers
Without built-in tools, a search-backed chatbot means writing the loop yourself: decide whether to search, call a search API, fetch pages, then hand the results back to the model. With the new tools you declare them in the request and the service runs the searches and page reads. Your client only renders the returned blocks and passes them back unchanged on the next turn, according to DeepSeek's Build a Chatbot guide.
- Web search: declare the deepseek_web_search tool with thinking disabled; the model runs web or image searches and answers with citations.
- Deep thinking search: the same tool with thinking enabled; the model searches and reads pages across several reasoning steps.
- URL reading: the deepseek_read_url tool fetches links found in the user's message.
- File input: PDF, Office and image files are parsed into the conversation context.
- Safety check: a safety_check parameter screens input and output and returns a refusal stop reason when it finds a risk.
- Extras from the same guide: a conversation-title option, a user metadata marker (date, location, language), read-aloud via TTS and voice input via ASR.
All three formats are served at api.deepseek.com. For SDKs, DeepSeek's pricing page lists https://api.deepseek.com as the base URL for OpenAI-format clients and https://api.deepseek.com/anthropic for Anthropic-format clients; the raw Messages path is /anthropic/v1/messages. The guide's examples send the key in an x-api-key header for Messages and as Authorization: Bearer for Responses and Chat Completions. In the Responses format, thinking is set with reasoning.effort and the output cap is max_output_tokens.
What it costs
Tool calls are billed on top of normal token charges. From DeepSeek's pricing page on 2026-11-15:
- Search (text or image) and web fetch: $0.75 per 1,000 calls each, whether you call the APIs directly or the model uses them as tools.
- File parsing and OCR: $0.225 per 1,000 pages.
- Text to speech: $0.45 per 10,000 characters list price, $0.225 during a limited-time 50% discount.
- Speech to text: $0.116 per hour of audio list price, $0.058 during the discount.
- deepseek-flash tokens: $0.30 per million cache-miss input tokens and $1.20 per million output tokens at peak; half that off-peak.
Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays; everything else, including weekends, is billed at the off-peak rate. The page does not say when the speech discount ends.
Our worked example, using those list prices and assuming no cache hits: a peak-hour turn with 10,000 input tokens, 1,700 output tokens and two searches costs about $0.003 + $0.002 + $0.0015, or roughly $0.0065. Off-peak, the token part halves and the same turn is closer to $0.004. Multi-turn chats that re-send context will often hit the cache, which is billed at $0.006 per million input tokens at peak, so real turns can cost less. Transcribing 300 hours of audio a month costs $34.80 at list price, or $17.40 while the discount lasts. These are our estimates, not DeepSeek's. To see what a turn actually used, the guide's sample response reports a usage.server_tool_use object with counts of web search and fetch requests.
Things that catch people out
- Some sampling settings are ignored. When a request declares deepseek_web_search or deepseek_read_url, sets safety_check to high, or enables the title, the service decides temperature, top_p and several other parameters itself, and the values you send are dropped. The guide lists them per format.
- max_tokens is per step, not per turn. It limits each inference step inside a round, so a multi-step search answer can produce more output in total than the number you set. (In the Responses format the parameter is max_output_tokens.)
- Public apps carry compliance duties. DeepSeek says public-facing apps must enable the safety check, and developers remain responsible for content moderation, applicable model registration and AI-generated content labelling.
- Old model names still work but are rerouted. deepseek-v4-flash and deepseek-v4-flash-vision-exp are accepted, but those models are retired and requests are served by DeepSeek-V4.1-Flash-1017 at the Flash price.
- The demo is a demo. DeepSeek says chat-demo keeps your API key and conversations in your browser and is for demonstration only. Its README adds that messages, attachments and speech are still sent to the API service, and that deleting all chats does not delete files already uploaded there.
Try it yourself
Try it with your AI. Paste this into your coding assistant to add DeepSeek's built-in web search to an existing chat backend.
I want to add DeepSeek's built-in web search and URL reading to my chat backend in
[language / framework].
First, check: which request format my code already uses (Anthropic Messages, OpenAI
Responses, or Chat Completions), and where the API key is read from. Do not print or
hard-code the key; read it from the DEEPSEEK_API_KEY environment variable. Ask me
whether I want plain web search (thinking disabled) or deep thinking search (thinking
enabled).
Then:
1. Point the client at DeepSeek with model "deepseek-flash". For an SDK, use the base
URL https://api.deepseek.com/anthropic (Anthropic format) or https://api.deepseek.com
(OpenAI format). Send the key as x-api-key for Messages, or Authorization: Bearer for
Responses and Chat Completions.
2. Add the tools deepseek_web_search and deepseek_read_url. In Messages each tool needs
both "type" and "name" set to the tool name; in the other formats "type" is enough.
Set safety_check to {"level": "high"}.
3. Render the returned blocks: thinking, search steps with result cards, page reads, and
the answer text with citation links. Check DeepSeek's Model Content Safety guide for
how a refusal is marked in my format (in Messages it is stop_reason "refusal") and
show my own refusal message.
4. Pass the returned blocks back unchanged in the next turn.
5. Remove temperature, top_p and the other parameters DeepSeek lists as ignored for my
format, and treat max_tokens (max_output_tokens in Responses) as a per-step limit.
6. Remember each search or fetch is billed at $0.75 per 1,000 on top of tokens.
Show me the diff and wait for my confirmation before changing files. Do not install
packages without asking.FAQ
Which model supports the built-in tools? DeepSeek's announcement names V4.1-Flash, called deepseek-flash in the API.
Do I pay for searches on top of tokens? Yes. Search and fetch are billed at $0.75 per 1,000 calls each, in addition to token charges.
Can I use ASR, TTS or OCR without the chat model? Yes. DeepSeek lists Search, Fetch, OCR, ASR and TTS as standalone APIs.
Source: DeepSeek: Full-Suite App API Release (2026-11-13), DeepSeek API docs: Models & Pricing, DeepSeek API docs: Build a Chatbot, deepseek-ai/chat-demo on GitHub
This story first appeared in the claudefox daily briefing for November 15, 2026.