DeepSeek-V4.1-Flash-1017 Unifies Search & Scraping for AI-Native Agents

Building agents with web search and data extraction previously required stitching together multiple third-party services: Tavily or Apify for search, Jina Reader or Firecrawl for Markdown conversion. DeepSeek's release of DeepSeek-V4.1-Flash-1017 alongside its self-developed Search and Fetch APIs removes this fragmentation by providing a unified, AI-native stack under one API key and base URL.

Model Improvements in V4.1-Flash-1017

The updated model retains the deepseek-flash API name and pricing (¥0.02/M input tokens cache hit, ¥4/M output tokens) but addresses the previous version's token inefficiency. Official benchmarks show significant gains:

Terminal-Bench 4.0: 31.2 → 42.3

DeepSWE v1.1: 74.2 → 77.0 (surpassing Claude Opus 5 at 74.0 and GPT-5.6-Sol at 73.0)

LMSYS Code Arena WebDev: 1670 points, ranking first among open-source models ahead of Kimi K3 Max at one-tenth the cost

Cost per task on DeepSWE v1.1 dropped from ¥3–4 to ~¥1 across low/high/max reasoning modes, making the model ideal for high-frequency search planning and data cleaning.

Search API: POST https://api.deepseek.com/v1/tools/search

Designed explicitly for LLM and agent consumption, the search endpoint accepts: query (string) with native advanced syntax: site:github.com OR site:huggingface.co, -site: exclusions, lang:zh,en, after:2025-01-01,

before:2026-06-01

ori_question

(string) — the user's original question to preserve intent during query rewriting and reranking count (1–50) — number of results returned

Each result returns a three-tier text gradient: snippet (100–200 chars) for UI display summary (600–800 chars) curated for LLM consumption — often sufficient without fetching full text content (full Markdown) when need_content=true Quality signals include rerank_score (0–3, ≥1.5 = high quality) and authority (0–4, ≥3 = highly authoritative), enabling simple list-comprehension filtering.

Fetch API: POST https://api.deepseek.com/v1/tools/fetch

Given a URL, returns title and cleaned Markdown content, eliminating manual DOM parsing with BeautifulSoup or Playwright.

End-to-End Pipeline in ~40 Lines of Python

import os, requests
from openai import OpenAI

API_KEY = os.environ["DEEPSEEK_API_KEY"]
HEADERS = {"Authorization": f"Bearer {API_KEY}", "Content-Type": "application/json"}

def search_and_extract(question: str, query: str):
    # 1. Call self-developed Search API with advanced syntax and original question alignment
    search_resp = requests.post(
        "https://api.deepseek.com/v1/tools/search",
        headers=HEADERS,
        json={"query": query, "ori_question": question, "count": 10}
    )
    search_resp.raise_for_status()
    results = search_resp.json().get("results", [])

    # 2. Filter by official rerank_score >= 1.5 for high-value results
    good_hits = [r for r in results if r["rerank_score"] >= 1.5]
    if not good_hits:
        return "No high-quality results retrieved"

    # 3. Fetch Markdown full text of the top result
    fetch_resp = requests.post(
        "https://api.deepseek.com/v1/tools/fetch",
        headers=HEADERS,
        json={"url": good_hits[0]["url"]}
    )
    full_md = fetch_resp.json().get("content", "") if fetch_resp.ok else ""

    # 4. Assemble long summaries and headline full text for DeepSeek-V4.1-Flash-1017 extraction
    context = "

".join(
        f"[{i+1}] {r['title']} ({r['url']}, authority:{r['authority']})
Summary: {r['summary']}"
        for i, r in enumerate(good_hits[:5])
    )
    if full_md:
        context += f"

[Core Page Full Text]
{full_md[:8000]}"

    client = OpenAI(api_key=API_KEY, base_url="https://api.deepseek.com")
    resp = client.chat.completions.create(
        model="deepseek-flash",  # auto-routes to DeepSeek-V4.1-Flash-1017
        messages=[
            {"role": "system", "content": "Extract key facts from retrieved web materials to answer the question, citing source URLs."},
            {"role": "user", "content": f"Question: {question}

Retrieved Materials:
{context}"}
        ]
    )
    return resp.choices[0].message.content

Function Tool Integration for Autonomous Agents

Both /v1/tools/search and /v1/tools/fetch can be wrapped as Function Tools with strict: true JSON Schema. DeepSeek-V4.1-Flash-1017 supports strict tool calling in both thinking and non-thinking modes, allowing the model to autonomously decide when to rewrite queries, which results to fetch deeply, and when to stop — enabling a DeepResearch-style agent in roughly 100 lines of code.

Ecosystem Integration & Pricing

The Search and Fetch APIs are already wired into DeepSeek Harness (≥0.2.1-rc.1), Claude Code, and Codex via server-side tool calls. Pricing is ¥5 per 1,000 requests (¥0.005 per call), billed per request regardless of count (1–50 same price). Only successful HTTP 200 responses with non-empty results incur charges; zero-result, 403, 404, or timeout calls are free. Rate limits: Search 100 QPS, Fetch 25 QPS, independent of model concurrency quotas.

Crucially, server-side search/fetch calls triggered through DeepSeek Harness, Claude Code, or Codex are completely free until 2026-11-01 00:00:00.

Demonstrated Use Cases

The author showcases successful extractions across diverse sources:

WeChat official account articles

Academic paper search and download

X (Twitter) tweet search

Xiaohongshu (Little Red Book) content queries

YouTube video search

These examples confirm the APIs handle Chinese social platforms, code repositories, and multimedia sources without additional adapters.