DeepSeek App APIs: Stacked Meters and Lock-In

On November 13, 2026, DeepSeek app APIs added web search, file input, URL reading, and a safety check as parameters on existing chat APIs. Voice stayed on standalone speech endpoints. Tool use is billed on separate meters on top of tokens, and public operators still own review, registration, and labeling.

Updated November 14, 2026

What did the November 13 release ship?

DeepSeek’s API changelog exposes web search, file input, URL reading, and model safety as parameters on the Messages, Responses, and Chat Completions APIs. The service runs those tool calls and assembles context on deepseek-flash, the V4.1-Flash-1017 model. Voice is not on that parameter list.

With thinking off and deepseek_web_search enabled, the model runs web or image searches and answers with citations. With thinking on, it searches the web and reads links across steps. deepseek_read_url fetches pasted links. PDFs, Office files, and images arrive as plain text or parsed layout.

GPTOnline reports upload limits of 100 MB and 1,000 pages per file, 20 documents per request, and extracted text under 8 MB. It also reports a context window of up to one million tokens.

The October 17 changelog opened the in-house search API. GPTOnline, Tech360, and Emergent say Search and Fetch were already standalone after that update. November 13 completes the set with OCR, ASR, and TTS, plus the in-chat tools. Fetch returns a page; it is not the search engine. The changelog places voice with those standalone APIs, as ASR and TTS, not as a chat parameter.

An MIT-licensed browser demo can import exported DeepSeek chat history and keeps keys and conversations on the device. DeepSeek describes it as a reference prototype rather than a production app.

How do in-chat tools differ from standalone calls?

Enable the tools and DeepSeek executes search, fetches, and parsing, then merges results into context. Call Search, Fetch, OCR, ASR, or TTS yourself when call timing must stay in your code. Crypto Briefing says the bundle could shrink that workload considerably for smaller teams—an interpretation, not a measured study.

Only search is described as running on DeepSeek’s index. The October 17 changelog says the engine handles hundreds of millions of search requests every day, indexes hundreds of billions of pages, and filters SEO spam, GEO manipulation, and low-quality AIGC. Parsing, OCR, ASR, TTS, and the safety parameter are not tied to that index in these sources.

Safety is an in-request parameter, not a standalone endpoint. The changelog’s standalone list is Search, Fetch, OCR, ASR, and TTS, with Search and Fetch already open before November 13.

What do the tool meters add on top of tokens?

Tools bill on top of tokens, on different meters. Only Search and Fetch are priced per successful request. Parsing and OCR are per page, ASR per audio hour, and TTS per characters. A Pandaily-sourced report and GPTOnline list the same dollar rates. GPTOnline adds that failed Search and Fetch requests, and requests that return no results, are not billed.

GPTOnline lists deepseek-flash at $0.30 per million input on a cache miss, $0.006 on a cache hit, and $1.20 per million output, halved off peak. The Pandaily-sourced report matches the $0.30 and $1.20 rates, the off-peak half, and five yuan on the China site beside the $0.75 rate.

Deep thinking searches and reads links across reasoning steps, and each tool type has its own meter. These sources do not say those in-chat reads are billed as repeated Fetch API calls. Emergent reports a 50 percent launch discount on ASR and TTS through the end of the year. GPTOnline calls that voice cut temporary.

Who still owns safety and the product surface?

A safety parameter reviews model input and output. GPTOnline reports HTTP 400 if input fails and a refusal marker if output is risky. The changelog says public services must enable the check and fulfill content safety review, large model registration, and AI labeling. Operators still need their own moderation policies.

DeepSeek owns in-chat orchestration, citations, and a basic refusal. You still own registration, labeling, local rules, multi-step latency, and the cost forecast. Teams weighing an agent instead of a cloned chat can look at DeepSeek’s open-source super-agent prototype, a separate codebase.

• Enable safety checks before any public launch, and handle HTTP 400 and refusal markers.
• Budget Search, Fetch, pages, and audio separately from cache-hit and off-peak tokens.
• Use in-chat tools when DeepSeek should assemble context, and standalone APIs when your code must time each call.