OpenAI web search vs Parallel vs Exa vs Tavily in 2026

Run the same question through all four

Each API keeps its own request and its own JSON on SpecStream. This script sends one question in Parallel's, Exa's and Tavily's format and to the SpecStream Search API, with one key, and prints the result count, the seconds each call took and the first link.

import os
import time
import requests

KEY = os.environ["SPECSTREAM_API_KEY"]
BASE = "https://api.specstream.ai"
Q = "who maintains the requests library for Python"
BEARER = {"Authorization": f"Bearer {KEY}"}

calls = {
    "Parallel": lambda: requests.post(f"{BASE}/parallel/v1/search", headers={"x-api-key": KEY},
                                      json={"search_queries": [Q], "mode": "fast"}, timeout=30).json()["results"],
    "Exa": lambda: requests.post(f"{BASE}/exa/search", headers={"x-api-key": KEY},
                                 json={"query": Q, "numResults": 10}, timeout=30).json()["results"],
    "Tavily": lambda: requests.post(f"{BASE}/tavily/search", headers=BEARER,
                                    json={"query": Q, "max_results": 10}, timeout=30).json()["results"],
    "SpecStream": lambda: requests.get(f"{BASE}/v1/web-search", headers=BEARER,
                                       params={"q": Q, "speed": "fast"}, timeout=30).json()["data"]["sources"],
}

for name, call in calls.items():
    start = time.perf_counter()
    results = call()
    seconds = time.perf_counter() - start
    print(f"{name}: {len(results)} results in {seconds:.2f} s, first: {results[0]['url']}")

The four calls cost $0.004 together, from the $10 of credit a new account gets. Change Q to a question from your own logs and run it again.

Replace OpenAI's web search tool in an agent

In the OpenAI Agents SDK, remove WebSearchTool from the agent's tools and pass SpecStream in mcp_servers. The agent calls search_web and read_pages like its other tools, and the same code runs with any model the SDK supports.

import asyncio, os
from agents import Agent, Runner
from agents.mcp import MCPServerStreamableHttp

async def main():
    params = {"url": "https://api.specstream.ai/mcp",
              "headers": {"Authorization": f"Bearer {os.environ['SPECSTREAM_API_KEY']}"}}
    async with MCPServerStreamableHttp(name="SpecStream", params=params, cache_tools_list=True) as server:
        agent = Agent(name="Researcher", instructions="Search the web with SpecStream.", mcp_servers=[server])
        print((await Runner.run(agent, "Who is the CEO of Mistral AI?")).final_output)

asyncio.run(main())

Code that calls a model directly defines a function tool that calls GET https://api.specstream.ai/v1/web-search, and passes the reply's data to the model as the tool result. The Agents SDK guide and the framework docs have both.

What 100,000 searches cost on each

Take an agent that makes 100,000 searches a month of 10 results each, at list prices.

SpecStream: $100 on any of the three paths or with speed=fast, and $500 on the default search of /v1/web-search, which adds the answer and rows. No tokens on top.

OpenAI web search: $1,000 in tool calls, then the search content tokens at the model's rates.

Parallel: $100 on turbo or fast, $500 on basic or advanced.

Exa: $400 on lite, $700 on auto.

Tavily: $800 for basic searches pay as you go, or $500 a month on the Growth plan of 100,000 credits. Advanced searches take 2 credits each.

The most accurate web search API in the world

The SpecStream Search API scores a 97.9 F-score on all 1,000 official SimpleQA Verified questions, with 978 correct. On WideSearch it scores 64.0 item F1 and 43.4 row F1 on all 200 official tasks, and the Extract API scores 0.908 F1 on WCXB. The scores, the method and the other APIs' published scores are at specstream.ai/benchmarks.

The Parallel, Exa and Tavily figures in the table come from Parallel's own run of September 2026. OpenAI's built-in search was not in that run.

What each path reads and returns

Parallel: POST /parallel/v1/search reads search_queries, mode, advanced_settings.max_results up to 100, source_policy and excerpt_settings, and returns results[] with url, title, publish_date and excerpts. The search runs on search_queries; objective is accepted and does not change it. The parallel-web SDK takes base_url="https://api.specstream.ai/parallel".

Exa: POST /exa/search reads query, numResults up to 100, includeDomains, excludeDomains, the published date range, includeText, excludeText and contents.text, and returns results[] with id, url, title, publishedDate and text. In exa-py: Exa(api_key=<key>, base_url="https://api.specstream.ai/exa").

Tavily: POST /tavily/search reads query, max_results up to 20, search_depth, include_raw_content, include_answer, the domain lists, the time range and topic, and returns results[] with title, url and content. In tavily-python: TavilyClient(api_key=<key>, api_base_url="https://api.specstream.ai/tavily").

Every parameter and reply field, with what we measured, is in the Parallel, Exa and Tavily migration guides.

Zero retention on any key

Set zero_retention when you create a key, or later. A search made with that key keeps no request text and no answer after the reply. The charge keeps the account, the key, the request ID, the price and the time. Set eu_egress_only and the key's searches leave only from our server in Madrid.

Search, pages and rows on the same key

The same key reads up to 20 pages a call with the Extract API at $0.80 per 1,000 pages, a page with no text not charged, and runs Deep Search, which returns every match on the live web as structured rows at 5¢, 15¢ or $1 a row. The most a Deep Search run can cost is approved before it starts.

Perplexity Search API

Ranked web results with no model: $5.00 per 1,000 requests, $1.00 with Fast Search. SpecStream answers Perplexity's POST /search at https://api.specstream.ai/perplexity, at $0.001 a search.

Brave Search API

Web search at $5 per 1,000 requests. SpecStream answers Brave's GET /res/v1/web/search at https://api.specstream.ai, at $0.001 a search.

Claude's web search tool

Web search on the Claude API costs $10 per 1,000 searches, plus standard token costs for the content it returns. SpecStream's MCP server and Search API give a Claude agent the same tool at $1 to $5 per 1,000, with no tokens billed for the search.