How to Use DeepSeek’s New API Tools: Search, Documents and Voice (2026)

On 13 November 2026, DeepSeek did something its cheapest-model reputation did not predict: it turned its consumer app into a developer platform. The web search, file reading, link reading and voice features that made the DeepSeek app feel complete are now available through the DeepSeek API — either switched on inside a chat with the model, or called on their own as standalone services. This guide explains what each tool does, what it costs, and which one to try first, using DeepSeek’s own changelog as the source of truth.

Quick answer: DeepSeek opened its search, file, URL and voice tools to developers on 13 November 2026. Here is what each tool does, what it costs, and which to try first.

The short version: if you already call the DeepSeek API, you do not need a new endpoint. You set the model to deepseek-flash, switch on the tools you want, and DeepSeek’s servers handle the tool calls and the context. If you would rather run the pieces yourself, five standalone APIs — Search, Fetch, OCR, speech-to-text and text-to-speech — let your code decide when each one runs.

What DeepSeek actually opened up on 13 November

DeepSeek’s own changelog entry, dated 13 November 2026, puts it plainly:

“In the DeepSeek App (including the web version), web search, file input, URL reading, voice interaction and safety work together with model inference to form the conversation experience users know. Today we open these capabilities to developers: enable the corresponding parameters in the existing APIs to get conversation features such as web search, file input and URL reading, with the tool calls and context assembly handled by the service.” — DeepSeek API Docs, Change Log, 13 November 2026

Two things are new, and one is not. New: the app’s features as parameters inside the existing Messages, Responses and Chat Completions APIs, running on the deepseek-flash model (the DeepSeek-V4.1-Flash-1017 checkpoint). New: standalone OCR, speech-to-text (ASR) and text-to-speech (TTS) APIs, plus an open-source demo chat client that rebuilds the app experience and can import your DeepSeek chat history. Not new: the standalone Search and Fetch APIs, which DeepSeek opened on 17 October 2026 — the 13 November change folds them into one complete suite rather than launching them.

The model behind all of it is deepseek-flash, currently the V4.1-Flash-1017 checkpoint. DeepSeek says the API calling method and pricing for the model itself are unchanged: you still call deepseek-flash to get the latest version. Our earlier coverage of SemiAnalysis GDPx put DeepSeek 4.1 Flash at 5.4% on 206 real professional tasks — cheap, but a long way from the frontier on hard work — so treat this release as a capability opening, not a capability leap: the best AI in the world scores just 12% on real work, and DeepSeek’s own benchmark scores in the changelog are DeepSeek’s own numbers, not independent results.

The five tools, and what each one is for

Tool What it does How you call it Try it first for
Web search (deepseek_web_search) The model runs web or image searches as the question requires and answers with citations Enable the tool with thinking disabled A support bot that must cite live prices, dates or docs
Deep thinking search With thinking enabled, the model searches the web and reads links on its own over several reasoning steps Enable the tool with thinking enabled Multi-hop research: “compare these three products, then check each claim”
File input Upload PDFs, Office files and images; text is extracted or the layout parsed, then added to the conversation Upload, then ask from the file Question-answering over a manual, a contract or a slide deck
URL reading (deepseek_read_url) The model detects links in your message, fetches the pages and answers with them in context Enable the tool and paste a link “Summarise this article and tell me what it leaves out”
Model safety Inputs and outputs are reviewed; a refusal marker returns when a risk is found Enable the safety-check parameter Anything public-facing — DeepSeek says public services must enable it

Alongside those five, the standalone APIs are the same capabilities with your code in charge: Search (/v1/tools/search) runs a query through DeepSeek’s own search engine and returns structured results; Fetch (/v1/tools/fetch) pulls a page’s content back as Markdown; OCR (/v1/tools/ocr) turns images and documents into text, page by page; ASR (/v1/tools/asr) turns speech into text in 21 languages; and TTS (/v1/tools/tts) turns text into spoken audio in 10 languages, with two voices — Stella and Tide.

How to switch the tools on

The setup is deliberately boring, which is the point. There is no new SDK to learn and no separate app to install:

  1. Use your existing DeepSeek API key. The tools run inside the APIs you already call — Messages, Responses or Chat Completions. If you have never called the API, create a key in the DeepSeek platform first; the tools add nothing to that step.
  2. Set the model to deepseek-flash. That name now points at the V4.1-Flash-1017 checkpoint. The older deepseek-chat and deepseek-reasoner names were retired in July 2026, so do not build on them.
  3. Switch on web search for a cited answer. Enable the deepseek_web_search tool with thinking disabled and the model searches the web and answers with citations. Turn thinking on as well and it becomes deep thinking search — slower, more thorough, and better for questions that need several sources.
  4. Switch on URL reading for link work. Enable deepseek_read_url and paste a link in the message; the model fetches the page and answers from it instead of guessing what it says.
  5. Upload files for document work. Send a PDF, Word, PowerPoint or image file and ask a question about it. DeepSeek extracts plain text for text-heavy documents, or parses the layout where tables and structure matter.
  6. Switch the safety check on before anything goes public. DeepSeek’s docs say public-facing services must enable the safety parameter and meet obligations including content-safety review, model registration and AI labelling. Treat that as a launch requirement, not a nice-to-have.

If you want the standalone route instead, call the tool endpoints directly — Search for a query, Fetch for a page, OCR for a document, ASR for an audio clip, TTS for a voice reply — and keep the model out of the loop until you need it. That is the right shape when your code, not the model, should decide when a tool runs: a search box, a document pipeline, or a voice-notes app.

What it costs

Tool charges sit on top of token charges — the model still bills per token, and every tool call adds its own fee. The model itself costs $0.30 per million uncached input tokens and $1.20 per million output tokens at peak hours, with off-peak rates at half that, according to DeepSeek’s pricing page. The tool prices below are from that same pricing page, as reported on 13–14 November 2026; the changelog itself does not list them, so check the pricing page before you build, because they can change.

Tool Price (as reported 13–14 Nov 2026)
Search $0.75 per 1,000 successful requests
Fetch $0.75 per 1,000 requests
OCR $0.225 per 1,000 pages
ASR (speech to text) $0.116 per hour of audio; $0.058 for a limited time (50% off, no end date given)
TTS (text to speech) $0.45 per 10,000 characters; $0.225 for a limited time (50% off, no end date given)

A worked example makes the scale concrete. One thousand searches ($0.75), 500 pages of OCR ($0.11), one hour of transcription at the launch discount ($0.06) and 10,000 characters of speech at the launch discount ($0.23) come to about $1.15 before token charges. That is the honest shape of this release: the tools are cheap enough to experiment with freely, and the bill that matters at scale is still the model’s tokens plus the volume of tool calls, not any single fee. Note the two voice discounts are labelled limited-time with no end date published, so do not budget on the discounted rate lasting.

Three starter builds that use the tools properly

1. A cited search assistant. Web search plus deep thinking search, with thinking on. Ask it a question that needs three sources and read the citations before you trust the answer — the citations are the feature, and a search answer without them is just a guess with better manners.

2. A document question-answerer. File input over a PDF manual or a contract, then questions in plain language. Start with one document and ten questions you already know the answers to; if the tool cannot answer those, it will not answer the eleventh one correctly either.

3. A voice-notes pipeline. ASR to transcribe a meeting recording, then the model to summarise it, then TTS to read the summary back. Keep clips under the ASR limit of 120 seconds and 8 MB per file, and remember the standalone Search API does not do image search — use the in-conversation web search tool for that.

DeepSeek has also open-sourced a demo chat client for these APIs, MIT-licensed, which rebuilds the app experience and can import your DeepSeek chat history; early reports put it past 1,100 GitHub stars within a day. It is the fastest way to see the tools working together before you write any code — and, usefully, it stores keys and conversations only in the browser, according to the same reports.

The limits to know before you build

Our take: the try-first ladder

The original element of this piece is a ladder, not a list. Most coverage of this release recites the five tools; the useful question is the order to try them in, because each one answers a different “is this real?” test:

  1. Web search first. One question, citations on. If the citations do not check out, nothing downstream matters.
  2. File input second. One document you know well, ten questions you know the answers to. This is the test that catches a parser that skims.
  3. URL reading third. One link, one summary, one “what does it leave out?” follow-up. This is where a model that guesses instead of reading gets caught.
  4. Voice last. ASR then TTS, because the voice discounts are the most likely price to move — build on the discounted rate and you will re-price the feature within months.

Run the ladder in an afternoon for about a dollar of tool charges plus tokens, and you will know more about this release than any launch post can tell you. The verdict from the docs alone is narrower: this is a genuine opening — the app’s whole toolkit, one bill, no new endpoint — but it is an opening for builders who already live in the DeepSeek API, not a reason to move there. The model is still the cheapest frontier-adjacent option with the weakest independent scores, and the tools do not change that. What they change is how far a cheap model can reach: into live web pages, into your documents, and into voice. That reach is worth a dollar to test.

Hands-on caveat: this guide is desk-researched from DeepSeek’s own changelog and pricing page as reported on 13–14 November 2026, plus six independent reports. No API key was used and no endpoint was tested hands-on for this piece; the steps, limits and prices are as documented, and the ladder above is the test to run before you trust them in production.

Frequently asked questions

What did DeepSeek actually open up on 13 November 2026?

It opened the features behind its consumer app to developers through the API: web search, deep thinking search, file input, URL reading and a safety check inside the existing Messages, Responses and Chat Completions APIs, plus standalone Search, Fetch, OCR, speech-to-text and text-to-speech APIs, and an open-source demo chat client. The model behind them is deepseek-flash (DeepSeek-V4.1-Flash-1017).

Do I need a new API key or a new endpoint?

No. The in-conversation tools are parameters on the APIs you already call — Messages, Responses or Chat Completions — with the model set to deepseek-flash. The standalone tools have their own endpoints under /v1/tools/, but they use the same API key and the same billing account.

How much do the new DeepSeek API tools cost?

Search and Fetch cost $0.75 per 1,000 requests, OCR costs $0.225 per 1,000 pages, speech-to-text costs $0.116 per hour of audio and text-to-speech $0.45 per 10,000 characters — with both voice services at half price for a limited time, with no end date published. These are DeepSeek’s pricing-page figures as reported on 13–14 November 2026, charged on top of the model’s token prices, so check the pricing page before you build.

Can I use the tools without the model?

Yes — that is what the standalone APIs are for. Search, Fetch, OCR, ASR and TTS each run on their own endpoint, so your code can call them directly and only bring the model in when you need it. The standalone Search and Fetch APIs actually opened on 17 October 2026; the 13 November release added the in-conversation tools and the standalone OCR and voice APIs.

What are the limits I should know about?

Speech-to-text takes clips of 120 seconds or less and 8 MB or less, in WAV or Ogg Opus, across 21 languages. OCR handles files up to 100 MB and documents up to 1,000 pages. Text-to-speech caps at 10,000 characters per request, with two voices in ten languages. The standalone Search API allows 100 requests per second and Fetch allows 25; image search is not available through the standalone Search API.

Sources and methodology

DeepSeek API Docs — Change Log, 13 November 2026 (primary source: “Releasing the full suite of application APIs”)

AI News Way — “DeepSeek Opens Its App Tools to API Users” (13 November 2026)

WebProNews — “DeepSeek Opens Its App’s Search, Files and Voice to Developers” (14 November 2026)

Crypto Briefing — “DeepSeek opens its full app toolkit to developers through its API” (14 November 2026)

AI Affairs — “DeepSeek opens its app’s search, file and voice tools through its API” (14 November 2026)

OpenAIMaster.ai — “The Best AI in the World Scores Just 12% on Real Work” (4 November 2026)

OpenAIMaster.ai — “AMD Says Its MI400 Beats Nvidia’s Vera Rubin by 30%” (14 November 2026)

OpenAIMaster.ai — “Gemini 4 Carbon Explained”

Method: every feature, tool name, endpoint and limit in this guide was checked against DeepSeek’s own API changelog entry dated 13 November 2026, read in full this run, and cross-checked against six independent reports published 13–14 November 2026. Prices are DeepSeek’s pricing-page figures as reported in that coverage; the changelog itself does not list tool prices. Benchmark figures for the V4.1-Flash-1017 checkpoint are DeepSeek’s own. Verification date: 15 November 2026. No endpoint was tested hands-on for this piece.

Leave a Reply

Your email address will not be published. Required fields are marked *