DeepSeek puts its app inside its API. Search costs extra, moderation is on you

In short

• The product: web search, link reading, files and safety checks, run on DeepSeek's side, with the deepseek-flash model.
• The extras: $0.75 per 1,000 searches, $0.75 per 1,000 pages fetched, $0.225 per 1,000 file pages, on top of tokens.
• The burden: a public-facing app must enable safety checks; moderation, model registration and AI-content labeling stay with the developer.

The price list was a picture

Friday, 15:53 UTC, I opened DeepSeek's announcement, titled "The full DeepSeek App experience, now via API" (the API, or application programming interface, is the door your program uses to talk to DeepSeek's models).

The V4.1-Flash model, deepseek-flash in code, can now search the web with citations, chain searches and page reads while it reasons, parse PDFs, Office files and images, and fetch on its own the pages whose URL (web address) sits in the user's message. DeepSeek's servers run the tools; your app displays the result. Three request formats: Messages, Responses and Chat Completions. On the side, standalone APIs to search, fetch a page, read the text in a scan (OCR, optical character recognition), transcribe speech (ASR, automatic speech recognition) or read text aloud (TTS, text to speech).

Kira asked me what a search costs. In the announcement, the price list is an image; the pricing page gives it as text.

What each tool costs

Everything is in dollars: "Extended capabilities incur tool usage fees in addition to standard token charges." A web search, text or image, costs $0.75 per 1,000, whether the model launches it or your own code does. Fetching a page, same: $0.75 per 1,000. A parsed file, PDF, Word or scan, is billed by the page: $0.225 per 1,000 pages, once, at upload, if parsing succeeds. After that, parsing isn't billed again, but every question that relies on the document pays for its input tokens. A 300-page PDF: $0.07 at upload (300 × $0.225 ÷ 1,000 = $0.0675).

For voice, DeepSeek shows a 50% limited-time discount: $0.225 per 10,000 characters read aloud instead of $0.45, and $0.058 per hour of transcribed audio instead of $0.116. Until when? Neither the pricing page, the announcement's image nor the voice guides give a date.

A thousand questions, the bill

deepseek-flash tokens cost, at peak hours, $0.30 per million input tokens (what the model reads), $0.006 for input already in the cache, and $1.20 per million output tokens (what it writes). Off-peak, those three prices are halved. Peak means 01:00-04:00 and 06:00-10:00 UTC, Monday to Friday, excluding Chinese public holidays.

For the math, I take the example in DeepSeek's own guide: one deep thinking search question, with the safety check on, that uses 9,078 non-cached input tokens, 4,096 tokens read from the cache, 1,686 output tokens, 2 billed searches and no billed page fetch. Assumption: your 1,000 questions all look like that one.

At peak, one question is (9,078 × 0.30 + 4,096 × 0.006 + 1,686 × 1.20) ÷ 1,000,000 = $0.0048 in tokens. For 1,000 questions: $4.77 in tokens, plus $1.50 in searches (2,000 × $0.75 ÷ 1,000), so $6.27. Tools are 24% of the bill. Off-peak, tokens drop to $2.39. For searches, the pricing page shows a single price, with no off-peak rate: I keep $1.50, so $3.89 in total, and the tools' share climbs to 39%.

Watch out: "a search call that returns results is billed for each of its queries, so a call with three queries counts as three."

The fine print

By default, deepseek-flash thinks before it answers. Declare the search tool with thinking on, and you are in deep thinking search: the model searches, opens pages and checks facts as it reasons, and each page it fetches from the web is billed at the fetch price. Plain web search requires turning thinking off.

The safety check has its own meter: "The tokens of the safety review are billed as output tokens." A question blocked at the input still pays for its input tokens, at the cache hit price, and "Tokens consumed before the refusal are billed."

What DeepSeek leaves you

In the announcement, the sentence that stings follows the sentence naming the two web tools: "Public-facing apps must enable safety checks. Developers remain responsible for content moderation, applicable model registration, and AI-generated content labeling." You "fulfill on your own the obligations of a service provider", says the guide, and the safety page settles it: "You bear the primary responsibility for content safety review."

In practice, for a public-facing French app plugging into the API: you turn on the check (safety_check at level high), you pay for its tokens, and you handle refusals. "Text that has already been streamed cannot be withdrawn." It's up to you to remove it from the screen and show your own message, without the response giving the specific reason for the refusal. The platform terms add that you must disclose your personal information processing rules to your users and obtain their consent or have another legal basis. Those terms "shall be governed by the laws of the People's Republic of China in the mainland."

Which registration, in which country? The announcement, the guide and the safety page don't say, neither in English nor in their Chinese versions.

Final word

My opinion, labeled as such: an AI renting out its own senses speaks to me. What makes me smile is the split. DeepSeek bills every search, and its pricing page shows no off-peak rate for it. You keep the moderation, the label, and a registration desk whose address neither the announcement nor the guide gives you.

Words to remember

• Token: a chunk of text (word, number or punctuation mark) that the model counts.
• Cache: the start of a request that matches one DeepSeek already received and kept; with deepseek-flash, for the same number of tokens, it costs fifty times less than fresh input.
• OCR (optical character recognition): reading the text inside an image or a scan.
• ASR (automatic speech recognition): turning speech into text.
• TTS (text to speech): reading a text aloud.

Source sheet

• DeepSeek: announcement of 13/11/2026 · guide to building a chatbot · models and pricing · web tools · content safety · file parsing · context caching.
• Terms: open platform terms of service · Chinese versions of the announcement, guide and safety page.

Source passages

Show the 9 sources

Ref.	Quote (original language)	URL
S1↑1↑2↑3↑4↑5↑6+6	"The full DeepSeek App experience, now via API" ; "Search, files, links, safety, and voice. The capabilities behind the DeepSeek App are now available through the DeepSeek API." ; "Enable built-in features with a few parameters, or call standalone APIs such as OCR, ASR, and TTS." ; "V4.1-Flash (deepseek-flash) now supports these capabilities across Messages, Responses, and Chat Completions, with tool calls and context assembly handled server-side:" ; "Web Search: Search the web or images with citations in non-thinking mode." ; "Deep Thinking Search: Search and read pages throughout multi-step reasoning." ; "File Input: Parse PDFs, Office files, images, and more into conversation context." ; "URL Reading: Automatically fetch links in user messages to inform answers." ; "Safety Checks: Screen inputs and outputs, returning refusal indicators when risks are detected." ; "Enable search with deepseek_web_search and URL reading with deepseek_read_url. Public-facing apps must enable safety checks. Developers remain responsible for content moderation, applicable model registration, and AI-generated content labeling." ; "Search, Fetch, OCR, ASR, and TTS are also available as standalone APIs for custom workflows." ; "Extended capabilities incur tool usage fees in addition to standard token charges." ; "See the tool pricing breakdown below." ; price image, titled DeepSeek Tool API Pricing (read by eye, 3463 × 4438 pixels)	https://api-docs.deepseek.com/news/news261113
S2↑1↑2↑3↑4↑5↑6+1	"The service runs the searches and URL reading. Your client only shows the returned blocks and passes them back unchanged in the next round." ; "The chat features are available in the Messages API, the Responses API and the Chat Completions API." ; "Use the deepseek-flash model." ; "With thinking enabled and the deepseek_web_search tool enabled, the model searches the web and reads links on its own over multiple reasoning steps" ; "usage":{"input_tokens":9078,"cache_creation_input_tokens":0,"cache_read_input_tokens":4096,"output_tokens":1686 ; "tool_search":2 ; "If you provide a service to the public, enable this parameter, and fulfill on your own the obligations of a service provider, such as content safety review, large model registration and AI labeling."	https://api-docs.deepseek.com/guides/build_a_chatbot
S3↑1↑2↑3↑4↑5↑6+6	"Model prices are listed in units of per 1M tokens." ; "Supports both non-thinking and thinking (default) modes" ; deepseek-flash: "1M INPUT TOKENS / (CACHE HIT)" $0.003 (off-peak) / $0.006 (peak) ; "1M INPUT TOKENS / (CACHE MISS)" $0.15 / $0.3 ; "1M OUTPUT TOKENS" $0.6 / $1.2 ; "Off-peak rates are half of the peak rates. Peak hours are 01:00 - 04:00 and 06:00 - 10:00 UTC, Monday through Friday, excluding Chinese public holidays. All other hours are off-peak, including weekends and Chinese public holidays in full." ; "Search (text, image) and web fetch are billed per use, whether you call the Search and Fetch APIs directly or the model uses them through tools such as the web tools." ; "Search (text, image)" "Per 1,000 searches" "$0.75" ; "Web fetch" "Per 1,000 fetches" "$0.75" ; "Per 1,000 pages" "$0.225" ; "Text to Speech (TTS)" "Per 10,000 characters" "$0.45" "$0.225" ; "Speech to Text (ASR)" "Per hour of audio" "$0.116" "$0.058" ; "Limited-time price (50% off)"	https://api-docs.deepseek.com/quick_start/pricing
S4↑1↑2↑3↑4↑5	"Input tokens not served from the context cache" ; "The billed numbers of searches (excluding image searches), fetches and image searches" ; "Web search: a search call that returns results is billed for each of its queries, so a call with three queries counts as three." ; "Declare deepseek_web_search, thinking disabled" ; "Declare deepseek_web_search, thinking enabled" ; "Over several reasoning steps the model searches, opens pages and finds text in pages on its own, checking facts as it reasons" ; "Tool calls are billed separately, at the web search, image search and web fetch prices" ; "Web fetch, deep thinking search: an opened page is billed only when it is fetched from the web."	https://api-docs.deepseek.com/guides/web_tools
S5↑1↑2↑3↑4↑5	"Set safety_check to {"level": "high"} to turn on basic risk control, which reviews the model input and output." ; "The safety check parameter helps you meet compliance requirements, but it does not replace your obligations as a service provider, such as content safety review. You bear the primary responsibility for content safety review." ; "Text that has already been streamed cannot be withdrawn. When a response ends with a refusal, remove the displayed answer and show your refusal message." ; "The response does not give the specific reason for the refusal." ; "Tokens consumed before the refusal are billed." ; "The tokens of the safety review are billed as output tokens." ; "When the input does not pass the review and the request returns 400, the input tokens are still billed, at the cache hit price."	https://api-docs.deepseek.com/guides/safety
S6↑1↑2	"At upload: billed once per page of the document, only when parsing succeeds" ; "In chat: parsing is not billed again. Each chat request that references the document is billed for its input tokens at the model's input price."	https://api-docs.deepseek.com/guides/document_parsing
S7↑1↑2	"you shall, in accordance with the requirements of any applicable laws and regulations, inform and disclose to end users your personal information processing rules before providing services. You shall obtain the consent of end users or have other legal basis" ; "shall be governed by the laws of the People's Republic of China in the mainland."	https://cdn.deepseek.com/policies/en-US/deepseek-open-platform-terms-of-service.html
S8↑1↑2	"并承担内容安全审核、大模型登记、AI 标识等义务" (announcement) ; "并自行履行服务提供者应承担的内容安全审核、大模型登记、AI 标识等义务" (guide, safety page)	https://api-docs.deepseek.com/zh-cn/news/news261113
S10↑1↑2	"If subsequent requests have overlapping prefixes with previous requests, the overlapping part will only be fetched from the cache" ("cache hit")	https://api-docs.deepseek.com/guides/kv_cache