When budgeting for a research assistant, count both the searches and the pages opened from the results.
DeepSeek’s V4.1-Flash API can now search the web, read URLs, and accept files. For a founder building a research assistant, the choice is whether to have the model call those tools or let the application decide what to retrieve. DeepSeek supports both paths, and the number of calls matters to the bill.
What DeepSeek added on November 13
DeepSeek’s release documents built-in web search, file input, URL reading, and safety checks for V4.1-Flash. The capabilities are available through its Messages, Responses, and Chat Completions APIs. DeepSeek also offers separate Search, Fetch, OCR, speech recognition and speech synthesis endpoints. This lets developers call, for example, OCR on a scanned document before passing selected text to a model. DeepSeek announcement
The official demo repository is a starting point for seeing how the App tools fit into a chat workflow. For a production integration, use the API reference for the interface you already support and test the response shape for each enabled tool. In particular, check what happens when a page fails to fetch, or a document returns partial text.
What 10,000 searches plus 10,000 fetches cost
DeepSeek lists web search and URL fetch at $0.75 per 1,000 uses each. At those rates, 10,000 searches cost $7.50 and 10,000 fetches cost another $7.50. The total is $15 in tool charges before model tokens, additional calls, or retries. DeepSeek lists file parsing and OCR at $0.225 per 1,000 pages. DeepSeek pricing
If a workflow searches once and opens three results, run the estimate with three fetches per search. That gives a tool budget of $30 for 10,000 such runs at the listed rates. The estimate assumes all four tool calls are billable and each run completes without a retry. Track actual calls per completed answer to see how closely your workflow matches the estimate.
A pilot for a small research assistant
Pick a question with an answer in a known public document and repeat it with a broken URL, a scanned PDF, and two sources that disagree. Record tool calls, the page or file each answer came from, and whether a reviewer can reproduce the cited result. That gives you a way to compare a model-directed run with a workflow that fetches an approved source list itself.
Start with a fixed tool-call budget per task. If the answer requires more retrieval, let the user resume with another budget. For an app that handles customer files, test the upload path with synthetic documents and decide which inputs your product may send to the API before you enable it for customers.
Abdessalam Alaoui
I'm Abdessalam Alaoui, founder of Musthave.AI. I write practical guides to AI models, tools, generative media.
