Until November 13, calling DeepSeek's API bought you a model and nothing else. It could not open a web page, could not read the PDF you had just uploaded, and had no idea what day it was. Anyone building something that felt like the DeepSeek App had to write the plumbing by hand: decide whether a question needed a search, call a search endpoint, fetch the pages, truncate them, glue the text back into the message array, then ask the model again. Five moving parts, five places to fail.
DeepSeek moved that whole chain server-side this week.
What actually changed
The model deepseek-flash — the API name for DeepSeek-V4.1-Flash — now takes a tools array that names capabilities rather than functions:
deepseek_web_search lets the model decide whether to search at all, and returns answers with source links.
Pair it with thinking mode and you get deep-thinking search: several rounds of searching and page reading inside a single reasoning pass.
deepseek_read_url fetches a link the user pasted and folds the page into context.
File input accepts PDF, Word, PowerPoint and images, parsed server-side, so nothing has to be converted before upload.
safety_check screens input and output and returns a refusal marker when something trips.
Both the OpenAI-style /v1/chat/completions route and the Anthropic-style /v1/messages route accept the same tools. An existing SDK integration needs a new base URL, a new model name and one extra line in the request body.

What it costs
Search and fetch are metered apart from tokens: $0.75 per thousand calls each, whether you use the tools or call the standalone endpoints. File parsing and OCR are $0.225 per thousand pages. Text to speech lists at $0.45 per ten thousand characters and is currently half that; speech to text is $0.116 per audio hour, also halved for now. Token prices did not move: for deepseek-flash, $0.15 per million input tokens on a cache miss off-peak, $0.30 at peak, and $0.60 to $1.20 per million output tokens.
The date problem nobody warns you about
This is the failure that fills every thread about the new API. Ask "what did DeepSeek ship today" and the model comes back with news from February. It is not broken. It has no clock, so it maps "today" onto whatever period was densest in its training data.
The fix is a system line carrying the date, which DeepSeek's documentation calls the latest reminder:
{"role":"system","content":"Today is 2026-11-17, Tuesday. Location: Shanghai. Interface: web."}
One community test went from twelve searches returning stale pages to three searches hitting the same-day announcement after adding that line. Regenerate it on every request. A cached system message is stale by tomorrow morning, and stale is the exact failure you were trying to prevent.
A minimal call
from openai import OpenAI
client = OpenAI(api_key="...", base_url="https://api.deepseek.com/v1")
r = client.chat.completions.create(
model="deepseek-flash",
messages=[
{"role": "system", "content": "Today is 2026-11-17, Tuesday."},
{"role": "user", "content": "What did DeepSeek release this week?"},
],
tools=[{"type": "deepseek_web_search"}],
)
print(r.choices[0].message.content)
Nothing else is required. No search key, no scraper, no chunking logic.
When the standalone endpoints are the better buy
If you do not need a conversation, do not pay for one. /v1/search returns result lists, /v1/fetch returns page text, /v1/ocr reads a scan, /v1/asr and /v1/tts handle audio. They are stateless: one call in, one result out. Batch invoice scanning belongs on /v1/ocr, not on a chat completion that will also bill you for thinking about it.
Three checks before you switch
Does your workload need search? Every triggered search is a metered call, and a model that searches too eagerly spends money quietly. Does your prompt carry a date? Add one before you blame the results. And are you on DeepSeek Harness 0.2.1-rc.1 or newer? Older harness builds do not match the current tool schema, and the failure looks like a model problem rather than a version problem.
The strategic read is simple. DeepSeek spent two years selling tokens. It is now selling the parts of the App people actually liked. For anyone building on top, that removes a week of plumbing and adds a line to the bill.