DeepSeek Turns Its Chat App Into an API: Search, Files, Voice, Priced

DeepSeek's chat app is free to use. On November 13, 2026, the company handed the machinery behind that app to developers. Web search, file parsing, URL reading, safety checks, OCR, speech to text and text to speech are now available through the DeepSeek API, either as built-in tools inside a chat call or as standalone endpoints you call directly.

The release matters because of how it is priced. Every tool carries its own per-use fee on top of the model tokens, and DeepSeek published the full rate card the same day. Search costs $0.75 per 1,000 queries. File parsing and OCR cost $0.225 per 1,000 pages. The two voice services are half price for a limited time, with no end date published. This article walks through what shipped, what each piece costs, what the limits are, and how the numbers compare with OpenAI.

What launched on November 13

DeepSeek's own changelog for November 13, 2026, is titled "Releasing the full suite of application APIs." The framing is deliberate. Inside the DeepSeek app, web search, file input, URL reading, voice interaction and safety work together with model inference to form the conversation experience users know. The release opens those same capabilities to developers, so a team can build a chatbot with the same behavior instead of wiring each piece itself.

There are two ways to use the new capabilities. The first is built-in tools: enable a parameter in an existing API call and DeepSeek handles the tool calls and context assembly on its servers. The second is standalone APIs: call Search, Fetch, OCR, speech to text (ASR) or text to speech (TTS) directly, for workflows where your own code decides when each tool runs.

One detail worth stating plainly: the standalone Search and Fetch APIs are not new. DeepSeek first announced its in-house search API on October 17, 2026, and the $0.75 search price was already live before this release. What is new on November 13 is the full suite arriving together, plus the built-in tools inside chat calls and the standalone OCR and voice services. Search pricing is unchanged; the scope of what you can buy is what grew.

The built-in tools inside a chat call

The built-in tools run on deepseek-flash, the API name for the DeepSeek-V4.1-Flash-1017 model, across the Messages, Responses and Chat Completions APIs. DeepSeek's documentation lists five of them.

The safety check deserves a closer read than the others. DeepSeek describes it as basic risk control, and its changelog is explicit about the boundary: public-facing apps must enable safety checks, and developers remain responsible for content moderation, applicable model registration and AI-generated content labeling. The tool flags risk; it does not take the compliance work off your plate.

DeepSeek also shipped two smaller conversation features with the same release: automatic conversation titles and user-context metadata, so a developer-built chatbot can behave more like the app, with titled threads and per-user context carried across turns. Its Build a Chatbot guide covers both, along with the web tools and file parsing.

The five standalone APIs

For pipelines where code, not the model, decides when a tool runs, DeepSeek split the capabilities into standalone endpoints. Each has its own billing unit, its own limits and its own endpoint.

Search and Fetch

The Search API runs searches through DeepSeek's internal search engine and returns structured results for search applications and AI agents. The Fetch API reads a URL you submit and returns its content. Both are billed per use at $0.75 per 1,000 calls, whether you call them directly or the model uses them through a built-in tool. Failed searches and empty results are not billed, according to launch coverage of DeepSeek's documentation.

OCR

The OCR API recognizes text in a PDF, Word or PowerPoint document or in an image, and returns the result page by page. It accepts PDF, Word (DOC, DOCX), PowerPoint (PPT, PPTX) and images such as PNG, JPEG, WebP, GIF, BMP and TIFF, and the format is detected from the file content rather than the extension. Two modes are offered: text recognition, which extracts text directly from Word and PowerPoint files without images and recognizes other files page by page, and rich layout recognition, which recognizes every file page by page with layout preserved.

The limits are strict and worth knowing before you build: a file may be at most 100 MB, a document may contain at most 1,000 pages, the upload must finish within 120 seconds, and the whole request must complete within 180 seconds. Files sent to the OCR API are not stored in the Files API and are not passed to a chat model. If you want the model to answer questions about a document instead of returning raw text, DeepSeek points you to file parsing inside a chat call, not the OCR endpoint.

Speech to text (ASR)

The ASR API transcribes an audio file into text, for example for voice input in a chatbot. It detects the language from the audio, with no language parameter to set, and supports 21 languages including Chinese, Cantonese, English, Japanese, Korean, German, French, Spanish, Portuguese, Italian, Russian, Arabic, Hindi, Indonesian, Vietnamese, Thai, Turkish, Malay, Filipino, Polish and Hungarian. The whole audio file is sent in one request and the complete transcript is returned; streaming recognition is not supported.

The format limits are narrow: 16-bit PCM WAV or Ogg Opus only, at most 8 MB, at most 120 seconds of audio, mono or stereo, with only the left channel recognized for stereo files. The upload must finish within 10 minutes. A 120-second cap means the endpoint fits voice notes and short clips, not meeting recordings or podcasts, and anything longer needs to be split client-side first.

Text to speech (TTS)

The TTS API converts text into natural speech, for example to read chatbot answers aloud. It offers two voices, Stella, female and the default, and Tide, male, across 10 languages: Mandarin Chinese, Cantonese, English, Japanese, Korean, Malay, Arabic, Indonesian, Thai and Vietnamese. The language is detected from the text, with no language parameter.

There are two modes. Full text takes the complete text in a JSON body and streams the audio back in the response. Streaming uses a WebSocket connection and sends text in chunks, for example while the model is still generating, with audio frames returned as they are synthesized. A request may contain at most 10,000 billed characters, the request body may be at most 64 KB, and one request returns at most 30 minutes of audio. Markdown formatting is removed before reading by default, which can be switched off.

What it actually costs

DeepSeek's pricing page lists the tool rates separately from the model token rates, and the split is the first thing to understand. Tool fees are charged in addition to token charges. Every search, fetch, parsed page or spoken character adds its own line to the bill, and the tokens the results add to the conversation are billed on top of that.

The model underneath is deepseek-flash, served by DeepSeek-V4.1-Flash-1017, with a 1 million-token context window and up to 384,000 output tokens. Its token rates depend on the time of day. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays. All other hours are off-peak, including weekends and Chinese public holidays in full, and off-peak rates are exactly half the peak rates.

A worked example makes the two-part billing concrete. Take a support chatbot that runs 10,000 web searches in a month, with 1 million input tokens and 500,000 output tokens, all off-peak. The searches cost 10 times $0.75, which is $7.50. The tokens cost $0.15 plus $0.30, which is $0.45. The month totals $7.95, and the searches are 94 percent of it. That is the shape of this release: the model is nearly free at these volumes, and the tools are the bill.

Voice shows the same pattern at a different scale. One hundred hours of transcription at the limited-time ASR rate costs 100 times $0.058, which is $5.80, against $11.60 at the list rate. One million characters of speech synthesis costs 100 times $0.225, which is $22.50 at the limited-time rate, against $45.00 at list. Ten thousand pages of OCR costs 10 times $0.225, which is $2.25. These are small numbers, and that is the point: the per-use rates are low enough that a side project can run on them, while the limited-time voice discount is the one line that can double without warning.

For readers outside the United States, the same rates convert to roughly £0.57 and C$1.07 per 1,000 searches, and roughly £5.68 and C$10.70 for the $7.50 in the example above, at November 15, 2026 market rates of 1 USD to 0.7569 GBP and 1.4271 CAD. Treat these as approximations. DeepSeek bills in US dollars, and the conversion is shown only so the numbers mean something outside the dollar.

The fine print that changes the math

Three details on DeepSeek's pricing page deserve attention before anyone budgets against the headline rates. First, the voice discount has no published end date. The pricing page labels both voice services as limited-time at 50 percent off, and DeepSeek has not said when the discount ends. One launch report described the discount as running through the end of the year, but DeepSeek's own page gives no date, so this article does not treat that as a fact. Budget voice work at the list rate and treat the discount as a bonus, not a baseline.

Second, DeepSeek reserves the right to adjust prices, a line carried on the pricing page itself. The rates in this article were checked against DeepSeek's own documentation on November 15, 2026, and they can move. A production budget needs a periodic check against the live page, not a screenshot of a launch article.

Third, the deduction order matters for anyone sitting on granted credit. Tool fees are deducted directly from the topped-up balance or the granted balance, with a preference for using the granted balance first, according to the pricing page. Granted credit is the balance DeepSeek gives out through promotions and programs, so free credit gets spent before paid credit when tool fees land.

DeepSeek vs OpenAI on search pricing

The sharpest comparison in this release is web search, because OpenAI publishes the same kind of line item. OpenAI's pricing page lists its web search tool at $10.00 per 1,000 calls, plus search content tokens billed at the model's rates. DeepSeek charges $0.75 per 1,000 searches, with the tokens the results add billed at deepseek-flash rates. On the call fee alone, DeepSeek's rate is about 13 times lower than OpenAI's.

The comparison needs its caveats, and they cut both ways. OpenAI's search runs on its own index with its own retrieval quality, and its content tokens are billed at the rate of whichever model you pair it with, which can be far more expensive than deepseek-flash. DeepSeek's search is cheaper per call, but no independent test has yet compared the two on answer quality or citation accuracy, and DeepSeek's retrieved content still costs tokens at deepseek-flash rates. The honest way to read the table is that DeepSeek is the cheaper call, not the proven better search. Run both on the same questions before choosing.

Google offers a different kind of comparison point. Its Gemini Decisions API charges $0.10 per million input tokens and bills no output at all, but it answers fixed questions with probabilities instead of generating text, so it is a decision tool rather than a search tool. The three vendors are pricing three different jobs: DeepSeek sells retrieval, OpenAI sells retrieval attached to expensive models, and Google sells scoring. Pick the job first, then the price.

How to try it

Getting started needs an API key from the DeepSeek platform and the standard base URL, https://api.deepseek.com, which accepts both OpenAI-format and Anthropic-format calls. The model name is deepseek-flash. The built-in tools are enabled with parameters in the existing Messages, Responses or Chat Completions calls: deepseek_web_search for search and deepseek_read_url for URL reading, with file input and safety checks enabled the same way.

The standalone endpoints sit under /v1/tools/: /v1/tools/asr for speech to text, /v1/tools/tts for text to speech, and /v1/tools/ocr for OCR, each with its own request format documented in DeepSeek's API guides. ASR takes a multipart file upload. TTS takes a JSON body for full text or a WebSocket connection for streaming. OCR takes a multipart upload with a file and a purpose field.

DeepSeek also published an open-source frontend demo for the release, at github.com/deepseek-ai/chat-demo, with support for importing DeepSeek conversation history, plus a hosted demo. That is a quick way to see the full app experience running in your own interface before writing any integration code: clone the demo, point it at your API key, and test search, file input and voice in one place.

One practical warning before you ship anything public. DeepSeek's changelog says public-facing apps must enable safety checks, and developers remain responsible for content moderation, model registration and AI-generated content labeling. The safety tool is basic risk control, not a compliance service. If your app faces the public, budget time for the moderation and labeling work the docs assign to you, because the API will not do it for you.

What to watch before you build on it

The release is two days old, and the honest gaps are the usual ones for a launch this fresh. No independent evaluation of the new tools has been published yet. The search quality, citation accuracy, OCR accuracy on messy scans, and transcription accuracy across the 21 ASR languages are all DeepSeek's own claims until someone else measures them. The benchmarks that exist cover the model, not the tools. For background on the serving stack behind the model, DeepSeek's open-source inference engine is covered in our DeepInfer explainer.

The built-in tools are listed for deepseek-flash only. DeepSeek's documentation does not list them for deepseek-v4-pro, the other model on the platform, so teams running the Pro model should not assume the app features carry over. The Pro model costs more per token, at $0.66 off-peak and $1.32 peak for cache-miss input, and its context window and feature set are documented separately.

The voice discount is the most likely line to move. Both voice services are half price for a limited time with no published end date, so any voice workload budgeted at $0.058 per hour or $0.225 per 10,000 characters should have a plan for the day those rates double. The search and OCR rates carry no discount label, so they are the more stable part of the card.

Finally, remember that this is a developer platform, not a consumer product. There is no paid DeepSeek app subscription behind this release; the app remains free and the paid side is the API. The November 13 release reads as DeepSeek turning its free app into a paid toolkit for developers, with the same features, the same model, and a per-use bill attached to each piece.

Common mistakes to avoid

Treating the tool prices as the whole bill is the first mistake. Tool fees sit on top of token charges, and a search-heavy workload will spend most of its money on searches, not on the model. Work the example in the pricing section with your own volumes before you commit to an architecture.

Sending long audio to the ASR endpoint is the second. The 120-second and 8 MB caps are hard limits, and the endpoint does not stream, so a meeting recording needs splitting and stitching on your side. If your audio is long, budget that work or use a different transcription service.

Using the OCR endpoint to ask questions about a document is the third. The OCR API returns recognized text page by page and never passes the file to a chat model. Questions about a document belong in file parsing inside a chat call, where the model reads the parsed content.

Assuming the voice discount is permanent is the fourth. It is labeled limited-time with no end date. Build the business case at list prices and let the discount be upside.

What did DeepSeek release on November 13, 2026?

DeepSeek opened the capabilities behind its chat app to developers: web search, deep thinking search, file input, URL reading and safety checks as built-in tools for deepseek-flash, plus standalone Search, Fetch, OCR, speech to text and text to speech APIs. An open-source frontend demo was published alongside the release.

How much does DeepSeek's web search API cost?

Search costs $0.75 per 1,000 searches and web fetch costs $0.75 per 1,000 fetches, whether called directly or used by the model through a built-in tool. The tokens the results add to the conversation are billed separately at deepseek-flash rates, and failed searches and empty results are not billed.

Is DeepSeek's API free?

No. The DeepSeek app is free, but the API is paid. Model tokens start at $0.15 per million input tokens off-peak for deepseek-flash, and every tool adds its own per-use fee.

What are the limits on DeepSeek's OCR, ASR and TTS APIs?

OCR accepts files up to 100 MB and 1,000 pages. Speech to text accepts at most 8 MB and 120 seconds of audio in 21 languages, with no streaming. Text to speech accepts at most 10,000 characters per request in 10 languages, with two voices, Stella and Tide, and returns at most 30 minutes of audio per request.

Can I use the new app tools with deepseek-v4-pro?

DeepSeek's documentation lists the built-in tools for deepseek-flash only and does not list them for deepseek-v4-pro. Use deepseek-flash, the API name for DeepSeek-V4.1-Flash-1017, for web search, file input, URL reading and safety checks.

How does DeepSeek's search price compare with OpenAI's?

DeepSeek charges $0.75 per 1,000 searches while OpenAI charges $10.00 per 1,000 web search calls, so DeepSeek's call fee is about 13 times lower. Both bill the retrieved content as input tokens on top, so the total depends on how much text each search returns.

DeepSeek's November 13 release is a straightforward deal: the company took the tools that make its free app useful, put a per-use price on each one, and handed developers the whole set. The rates are low, the limits are clearly documented, and the voice discount is the one line to watch. If you build chatbots, document pipelines or voice features, the suite is worth a test run this week, while the voice services are still half price and the documentation is fresh.

Sources

• DeepSeek API Docs, Change Log, November 13, 2026: Releasing the full suite of application APIs
• DeepSeek API Docs, Models and Pricing (checked November 15, 2026)
• DeepSeek API Docs, Build a Chatbot guide
• DeepSeek API Docs, Speech to Text guide
• DeepSeek API Docs, Text to Speech guide
• DeepSeek API Docs, OCR guide
• DeepSeek, chat-demo on GitHub (open-source frontend demo)
• OpenAI API Pricing (web search $10.00 per 1,000 calls, checked November 15, 2026)
• AI News Way, DeepSeek Opens Its App Tools to API Users, November 14, 2026
• AI Affairs, DeepSeek Opens Its App's Search, Files and Voice to Developers, November 14, 2026
• PromptWire, DeepSeek Application APIs: Files, Voice, Pricing and Limits, November 14, 2026
• WebProNews, DeepSeek Opens Its App's Search, Files and Voice to Developers, November 14, 2026