Web Search Capabilities of Major International LLM APIs

🕐 2025-07-20 ✍️ Moyi Foreign Trade

ChatGPT (OpenAI), Gemini (Google) and Claude (Anthropic) have officially built web search into their APIs, all provided as “server-side retrieval tools” — easy to call and automatically returning real-time content with citations.
DeepSeek-R1 offers an “enable_search” parameter, but it currently only works on Alibaba Cloud's LLM API gateway and in OpenSearch scenarios — a region/version-limited public beta feature.
ByteDance's Doubao and Tongyi Qianwen have not yet opened official search tools in their general chat-completions APIs; they can only be called via the Ark plugin (Doubao) or the “official website workflow/web toolchain” (Qianwen), while native SDK/HTTP endpoints still require integrating an external search API yourself.
Pricing differs significantly between vendors — OpenAI and Anthropic bill both “search calls” and “generation tokens”, Google prices search separately at “$35/1,000 retrievals”, and DeepSeek is temporarily free of separate charges on the Quark Search channel.
The most granular feature set is Claude: you can limit max_uses, whitelist/blacklist domains, localize user_location, and it supports multi-round progressive retrieval; Gemini lets you control whether to search via the dynamic retrieval_threshold; OpenAI offers the search_context_size option to balance quality and cost.

Side-by-Side Comparison of Web Search Capabilities

Model	Built-in search	How to enable / key parameters	Return format & citations	Billing	Main limitations & notes
DeepSeek-R1	Limited public beta	`enable_search:true` (only deepseek-r1, via Alibaba Cloud LLM API / OpenSearch)	Summary + URL	Priced with DeepSeek tokens for now; no separate search fee	Relies on Quark Search, usable in mainland China; other DeepSeek models not yet supported
Doubao (豆包)	No general API	Configure the “web search plugin” in Volcano Ark bots/applications; the API side receives structured JSON results	Assembled by yourself	Same price as inference tokens; plugin traffic free for 1,000 calls/day	Ark platform only; OpenAI-compatible / direct REST cannot use `enable_search` yet
Qwen (Tongyi Qianwen)	No general API	DashScope web “toolchain” includes search; API not opened	Summary + URL	N/A	Users must build their own RAG or use Bing/Brave APIs + function calling
ChatGPT (OpenAI)	✅	a) Chat Completions: pick `gpt-4o-search-preview` and pass `web_search_options`; b) Responses API: add `{"type":"web_search_preview"}` to the `tools` array	Inline citations via `url_citation`	Search fee included in model pricing ($22.5 per million tokens), with cost tiers by context_size7	4o series only; beta stage, some API parameters unsupported6
Gemini (Google)	✅	`tools:[{"google_search":{}}]`, usable in `client.models.generate_content` or Vertex AI Chat Completions	Inline links + `Search Suggestions`	$35 / 1,000 retrievals, tokens billed separately	Dynamic retrieval threshold `retrieval_threshold`; can mix with up to 10 retrieval sources
Claude 3.x	✅	`tools:[{"type":"web_search_20250305","name":"web_search","max_uses":N,...}]`	Forces source attribution; supports `allowed_domains` / `blocked_domains` / `user_location`	$10 / 1,000 retrievals + model tokens	Supported by Claude 3.7 Sonnet, 3.5 series and 4 Opus; iterative retrieval by default, limitable via `max_uses`

Model-by-Model Breakdown

1. DeepSeek

In the Alibaba Cloud LLM API console, you can enable the “web search” plugin for deepseek-r1. Set the following in the request body:

"enable_search": true

The backend then calls Quark Search, returns summary snippets and concatenates them into the context, appending short URL citations at the end of the message. The feature is still in public beta and has not been opened to other models like DeepSeek-V3/V2.

2. Doubao

Volcano Engine Ark supports attaching a “web search plugin” to bots/workflows, with the return structure defined in the official docs (fields such as site_name/summary/url). The plugin actually calls ByteDance's own search (including Douyin Encyclopedia, ByteSearch, etc.), but the OpenAI-compatible chat/completions endpoint still has no enable_search field — you need to go through Ark flows or integrate an external search source yourself.

3. Qwen (Tongyi Qianwen)

The DashScope web version has bundled Qianwen with “web search and page parsing”, but the SDK / HTTP API has not yet exposed the corresponding parameters. Developers typically implement retrieval via function calling plus Bing/Brave SERP APIs, then write results back into the context. Alibaba Cloud OpenSearch is currently only piloting this for DeepSeek-R1; Qianwen is not yet included.

4. ChatGPT / OpenAI

Two calling paths

Chat Completions: directly pick gpt-4o-search-preview / gpt-4o-mini-search-preview and attach web_search_options (search_context_size/user_location, etc.).
Responses API: any model + tools:[{"type":"web_search_preview"}] auto-retrieves when needed6.

Output: citations embedded inline via url_citation, with full URLs in annotations.
Pricing: retrieval uses a separate model with slightly higher token prices; different search_context_size values affect cost7.

5. Gemini

When calling the Gemini Developer API or Vertex AI, add:

"tools": [{"google_search":{}}]

This enables Grounding with Google Search: Google handles querying, deduplication and ranking in the background and returns answers with source links.

You can fine-tune whether to search via dynamic_retrieval.rerank_threshold (AI Studio UI) or the SDK's retrieval_threshold; user_location automatically affects localized results. Billing is per retrieval request at $35/1,000.

6. Claude

Add the "type":"web_search_20250305" tool to the tools field and specify max_uses; Claude decides the query terms itself, iterates multiple times and produces results with citations.
Advanced controls:

Domain filtering: allowed_domains / blocked_domains
Region: user_location
Iteration depth: max_uses — exceeding it triggers a max_uses_exceeded error1.
Cost: $10/1,000 retrievals + model tokens, now covering Claude 4 Opus, 3.7 Sonnet and 3.5 Haiku/Sonnet.

Selection Recommendations

Need “official-grade” real-time info right away: go with Gemini (quality + global retrieval coverage) or Claude (fine-grained citation control); if you're already in the OpenAI ecosystem, GPT-4o-search-preview is a quick way in.
Team is based in mainland China and needs filing/compliance: DeepSeek-R1's Quark Search channel is the currently available public-beta option; Doubao currently requires the Ark plugin; for Qianwen, combine it with self-built RAG.
Strictly restrict information sources or reduce hallucinations: Claude's domain allow/deny lists and max_uses give the strongest control.
Cost-sensitive with high query volume: run retrieval on your own backend and use SERP APIs + function calling to let any model read the search results, completely avoiding platform search fees.

Reference Code Snippets

# OpenAI Responses API calling built-in search
from openai import OpenAI
client = OpenAI()
resp = client.responses.create(
    model="gpt-4o",
    tools=[{"type":"web_search_preview", "search_context_size":"low"}],
    input="最新 GPT-4o 模型的价格是多少？"
)
print(resp.output_text)

# Gemini Developer API calling Google Search Grounding
from google import genai
cli = genai.Client()
resp = cli.models.generate_content(
    model="gemini-2.5-flash",
    contents="今年奥运会乒乓球男单冠军是谁？",
    config=genai.types.GenerateContentConfig(
        tools=[{"google_search":{}}],
    )
)
print(resp.text)

# Claude 3.7 Sonnet enabling web_search, limited to 2 queries
import anthropic, os, json
client = anthropic.Anthropic(api_key=os.getenv("ANTHROPIC_API_KEY"))
msg = client.messages.create(
    model="claude-3-7-sonnet-20250219",
    max_tokens=1024,
    tools=[{
       "type":"web_search_20250305",
       "name":"web_search",
       "max_uses":2,
       "allowed_domains":["bbc.com","reuters.com"]
    }],
    messages=[{"role":"user","content":"请给出最近一次英国基准利率决议要点"}]
)
print(json.dumps(msg,indent=2))

Summary: truly native, out-of-the-box “web search” currently exists only in OpenAI GPT-4o series, Google Gemini, Anthropic Claude and the small-scale public beta of DeepSeek-R1. Others (Doubao, Tongyi Qianwen, etc.) still need plugins or external search APIs. When choosing, weigh ease of calling, citation transparency, cost and regulatory compliance, and keep a fallback plan (self-built RAG or function calling) ready for service changes.

🌐 阅读中文版