OpenAI API Web Search Guide

Responses API + web_search tool gives you live web access with a default-group key. gpt-5.5 / gpt-5.4 verified with real search and cited sources, billing included.

This page explains how to use web search with GPT models on APIYI, verified by hands-on testing in June 2026.

TL;DR

APIYI fully supports OpenAI’s official web search: use the Responses API (/v1/responses) with the web_search tool. Both gpt-5.5 and gpt-5.4 were verified to genuinely search the web and return up-to-date information with source citations. A default-group key works out of the box — no special activation needed.

Endpoint:  POST https://api.apiyi.com/v1/responses
Tool:      tools: [{"type": "web_search"}]
Models:    gpt-5.5 / gpt-5.4 (verified)

Real-world availability (test data, 2026-06-11)

Model	Web result	Citations	Searches per Q&A	Latency
gpt-5.4	✅ Accurate same-week news	✅ Structured url_citation	1	~11s
gpt-5.5	✅ Accurate same-week news (auto time-window scoping, multi-source cross-checking)	✅ Structured url_citation	~8	~51s

Model choice: pick gpt-5.4 for speed and cost; pick gpt-5.5 for coverage and rigor (more search rounds and larger retrieved-content injection mean higher cost and latency — see the billing section).

Quick start

cURL

curl https://api.apiyi.com/v1/responses \
  -H "content-type: application/json" \
  -H "authorization: Bearer YOUR_APIYI_KEY" \
  -d '{
    "model": "gpt-5.4",
    "max_output_tokens": 8192,
    "tools": [{"type": "web_search"}],
    "input": "What new models has Anthropic released in the past week? Search and include source links."
  }'

Python (OpenAI SDK)

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_APIYI_KEY",         # default group works
    base_url="https://api.apiyi.com/v1",
)

resp = client.responses.create(
    model="gpt-5.4",                  # or gpt-5.5
    max_output_tokens=8192,           # recommend >=8k; gpt-5.5 uses many reasoning tokens, too small -> incomplete
    tools=[{"type": "web_search"}],
    input="What new models has Anthropic released in the past week? Search and include source links.",
)

# 1) Final answer text
print(resp.output_text)

# 2) Actual number of searches in this call (billing basis, see below)
search_calls = [item for item in resp.output if item.type == "web_search_call"]
print(f"Searches in this call: {len(search_calls)}")

# 3) Source citations (structured url_citation)
for item in resp.output:
    if item.type == "message":
        for content in item.content:
            for ann in getattr(content, "annotations", []) or []:
                print(f"Source: {ann.title} | {ann.url}")

Response structure

The output array contains, in execution order:

item type	Meaning
web_search_call	One actually executed search (billing counts these entries)
reasoning	The model’s reasoning process (gpt-5 series)
message	The final answer; its content[].annotations contains url_citation (title + url)

status: "completed" means it finished normally; incomplete usually means max_output_tokens was too small — increase it.

Billing (important)

Web search incurs a tool-call fee, made up of two parts:

Item	Price	Notes
Tool-call fee	$10 / 1,000 calls ($0.01 per call)	Tool name: web_search; counted by the number of web_search_call entries in the response output — one question may trigger multiple searches (gpt-5.4 usually 1, gpt-5.5 usually 5–8)
Retrieved-content token fee	Standard model input price	Search results are injected into the model context and billed as input tokens. This is usually the larger share: measured at roughly 9k input tokens per Q&A for gpt-5.4, and 48–54k for gpt-5.5

Measured total cost per web-enabled Q&A: gpt-5.4 ≈ $0.01 search fee + 9k tokens; gpt-5.5 ≈ $0.08 search fee + ~50k tokens. Estimate against your expected query volume.

Notes

1. Use the Responses API — do not use Chat Completions’ web_search_options: gpt-5 series models do not support that parameter (official OpenAI behavior; returns 400 Unknown parameter: 'web_search_options'). web_search_options only applies to the dedicated *-search-preview models.
2. Set max_output_tokens to at least 8192: gpt-5.5 consumes many reasoning tokens; a small limit returns status: "incomplete" with no final answer, while tokens are still billed.
3. The legacy tool type web_search_preview also works with identical behavior; for new integrations, use web_search directly.
4. To control cost, constrain search behavior in the prompt (e.g. “search at most 2 times”) or use gpt-5.4.

FAQ

Q: How do I confirm the answer really used the web?

A: Check whether the response output contains entries with type="web_search_call" and whether the message annotations include url_citation. Both present means real web access; answer text alone without these two markers means the model answered from training data.

Q: Do I need a different group or a special key?

A: No. For OpenAI models, a default-group key can call web search directly.

Q: Which models are supported?

A: gpt-5.5 and gpt-5.4 are verified. Other gpt-5 series models should also support the Responses API web_search tool in principle — run the verification check from the FAQ above before relying on it.

Related Docs

OpenAI Native Calls (Responses API)

Responses API endpoint, parameters, and setup

OpenAI Prompt Caching

The large input-token volume injected by web search pairs well with caching