Published: 13 November 2026
DeepSeek opened the features behind its chatbot app to developers on 13 November 2026. Through the API, its deepseek-flash model can now search the web and cite its sources, read PDFs, Office files and images, and open links in a message, each switched on with a single setting. DeepSeek also launched separate APIs for OCR, speech recognition and text to speech, with OCR priced at $0.225 per 1,000 pages.
Search, file reading and link reading are now settings on deepseek-flash
Until now, a developer who wanted a DeepSeek chatbot with live web answers had to add a search service, a page scraper and a PDF parser, and connect them to the model. DeepSeek now runs those steps on its own servers. You add the deepseek_web_search tool to a request, the model decides when to search, and it answers with citations. With thinking switched on, it searches in several rounds and makes up to about 20 tool calls before it answers.
Files go through the same API. You upload a PDF, a Word or PowerPoint file or an image, DeepSeek turns it into text, or into text plus cropped figures, and you refer to it in the chat. A second tool, deepseek_read_url, opens the first ten links in the user’s last message. A safety_check setting reviews input and output and returns a refusal when content fails the check, without saying why.
All of this works in the Messages, Responses and Chat Completions formats, so code written for Anthropic’s or OpenAI’s API needs little change. DeepSeek now also sells OCR, speech recognition and text to speech as separate APIs, next to the Search and Fetch APIs it opened in October.
OCR costs $0.225 per 1,000 pages, and Dutch mornings fall in peak hours
DeepSeek prices each part separately, according to its pricing page and API guides:
- deepseek-flash (DeepSeek-V4.1-Flash-1017, 1M token context): $0.15 per million input tokens and $0.60 per million output tokens off-peak, double at peak times
- Web search and link reading: $0.75 per 1,000 successful requests, and one search with three queries counts as three
- OCR and file parsing: $0.225 per 1,000 pages
- Speech recognition: $0.058 per hour of audio, a 50% launch discount on the list price of $0.116
- Text to speech: $0.225 per 10,000 characters, half the list price of $0.45
Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays. In Dutch winter time, the second window runs from 07:00 to 11:00, when most office work starts, so a European team pays the double rate for part of its working day. DeepSeek gives no end date for the speech discounts. The safety check has no separate price, but its review tokens are billed as output.
The speech APIs skip Dutch: 21 languages for listening, 10 for talking
Speech recognition handles 21 languages, including German, French, Polish and Hungarian, but not Dutch. Text to speech covers 10, mostly Asian languages plus English and Arabic. Neither API lets you choose a language, because it is detected from the audio or text. Speech recognition also takes clips of at most two minutes and does not stream, which makes it a poor fit for live calls. DeepSeek publishes no language list for OCR, file parsing or search.
DeepSeek’s privacy policy says personal data is stored and processed in China
DeepSeek’s privacy policy, last updated on 10 February 2026, says the company collects, processes and stores personal data in the People’s Republic of China. Its section for the European Economic Area repeats that. The new API guides name no other region and say nothing about the EU.
Two details matter for documents. Files you upload for parsing stay in your account until you delete them, unless you set an expiry of one hour to 30 days at upload. Files sent to the standalone OCR API are not stored, according to DeepSeek. China has no adequacy decision from the European Commission, so sending personal data there needs safeguards such as standard contractual clauses and a transfer impact assessment. DeepSeek’s policy names no transfer mechanism.
Public research is a fit, client files and Dutch voice agents are not
For the organisations DataNorth works with, the appeal is clear. Building search and document parsing around a model is often the slowest part of an AI assistant project, and DeepSeek now delivers both as a setting at a low price per call. That suits work on public information: market and competitor research, summaries of published reports, or reading tender documents and product manuals that are already public.
The documents most teams want to process are different: contracts, invoices, HR files and customer emails. Those contain personal data, and sending them to servers in China needs a decision by your privacy officer before a pilot, not after it. You also cannot limit DeepSeek’s search to your own website or trusted sources, because the API rejects domain filters, which is a real limit for a customer-facing bot. For Dutch speech, the new voice APIs are not an option at all.
Our advice: try the web search and OCR APIs this month on public material, and compare the cost per answer with your current setup. Keep personal and confidential files out until a data protection impact assessment (DPIA) and a transfer assessment are done. If you need the same features on sensitive data, an EU-hosted service or an open model running in your own environment is the safer route.
For more information, visit the official announcement of the full suite of application APIs in the DeepSeek API change log.