DeepSeek Turns Its Whole App Into an API: Web Search, Files, OCR and Voice Now Ship Inside deepseek-flash Calls

DeepSeek published a changelog entry on November 13 titled "Releasing the full suite of application APIs," and it does what the name promises: everything that makes the DeepSeek app feel like the DeepSeek app is now callable by developers. Web search, deep-thinking search, file input, URL reading and safety checks run inside regular model calls on deepseek-flash, the current version of DeepSeek-V4.1-Flash-1017, which carries a 1 million-token context window and up to 384,000 output tokens. The features work across all three API shapes — Messages, Responses and Chat Completions.

The mechanics are parameter switches rather than separate integrations. With thinking mode off, enabling the deepseek_web_search tool lets the model fire web or image searches as needed and answer with citations; with thinking on, it searches and reads links on its own across multi-step reasoning. The deepseek_read_url tool detects links a user pastes and fetches the pages into context. File input accepts PDF, Word, PowerPoint and image formats — up to 100MB and 1,000 pages per file, 20 documents per request, with total extracted text capped at 8MB. A safety_check parameter reviews inputs and outputs, rejecting failed inputs with an HTTP 400 error and returning a refusal marker for risky output. DeepSeek says public-facing services must enable it, alongside their own moderation and local compliance obligations.

The same capabilities are also broken out as standalone services: a Search API rated at up to 100 requests per second, a Fetch API that returns pages as Markdown at up to 25 requests per second, OCR for documents, speech recognition covering 21 languages, and text-to-speech in 10 languages with two voices — Stella and Tide — streaming over WebSocket. Tool calls and context assembly are handled server-side, which is the part developers previously had to stitch together themselves.

Pricing is published per tool. Search and web fetch cost 5 yuan (about $0.70) per 1,000 calls, OCR 1.5 yuan per 1,000 pages, speech recognition 0.39 yuan per hour and speech synthesis 1.5 yuan per 10,000 characters. The model itself remains at $0.30 per million input tokens and $1.20 per million output tokens before cache discounts. To prove the point, DeepSeek open-sourced a demo chat client built on the new APIs that can import a user's existing DeepSeek chat history — the company's own framing is that developers can now recreate the app one-to-one.

On capability, DeepSeek's October figures put deepseek-flash at 42.6 on HLE without tools and 77.0 on DeepSWE, with lower task runtime and output length than the previous version. Those are company-reported numbers and have not been independently reproduced.

Strategically, the release turns DeepSeek from a model supplier into a platform. Every major lab is racing to own the layer between the model and the user — Anthropic moved its managed agents onto customers' servers this week, and OpenAI put a wallet into ChatGPT — but DeepSeek's answer is the cheapest full stack around, and the only one that ships with a cloneable reference app. It also exports its home-market compliance frame: the safety parameter and the surrounding obligations follow the API wherever it is deployed.

The catch is scope. Everything here runs on deepseek-flash, the small fast model — not the flagship V4 Pro tier — so heavy reasoning and long agent runs still lack the new tools. The platform play is real, but the flagship is still waiting for its turn.