LLM web access API in 2026

Search and read on one key

A model needs two moves: find the pages, then read them. The Search API returns ranked sources with url, title, site, snippet and the published date where the page states it. passages=true adds the text of each page closest to the question, so most answers need no second call. When they do, the Extract API reads the URLs the model picked.

import os
import requests

H = {"Authorization": f"Bearer {os.environ['SPECSTREAM_API_KEY']}"}

found = requests.get(
    "https://api.specstream.ai/v1/web-search",
    params={"q": "EU AI Act obligations for general purpose models", "speed": "fast", "passages": "true"},
    headers=H,
    timeout=30,
).json()["data"]

context = "\n\n".join(
    f"{s['url']}\n{s.get('excerpt') or s['snippet']}" for s in found["sources"]
)

read = requests.post(
    "https://api.specstream.ai/v1/extract",
    headers=H,
    json={"urls": [s["url"] for s in found["sources"][:3]], "format": "markdown"},
    timeout=60,
).json()["data"]

for page in read["results"]:
    if page["status"] == "ok":
        print(page["title"], len(page["text"]))
print(found["price_usd"], read["price_usd"])

$1 per 1,000 searches

A fast search of up to 10 results costs $0.001. Each started block of 10 more adds $0.0012, so 50 results cost $0.0058. The default search, which also writes answer and rows, costs $0.005. A search that returns no results, a page with no text and a call answered 503 are not charged, and each reply carries its own price_usd.

Pages behind a refusal

A page that refuses a plain read is tried again as a browser would ask for it, then in a real browser, then from the site's own feed, then from its newest public archive capture. read_from says which one gave the text, and captured_at dates an archive copy. PDF, Word, Excel and EPUB files are read in the same call at the same $0.0008.

One MCP server for any client

Add https://api.specstream.ai/mcp (Streamable HTTP) to Claude, Cursor, VS Code or any MCP client, signed in with OAuth or an API key. The model gets search_web for a question and read_pages for up to 20 URLs, at the same prices as the API. Setup for each client is in the integration guides.

Keep the client you have

Code written for Tavily, Brave Search, You.com, Exa, Jina, Serper or the Google Custom Search JSON API keeps its request and its parser: change the base address to SpecStream's and send a SpecStream key. Each search through a drop-in address costs $1 per 1,000. The list with one curl line each is at https://api.specstream.ai/v1/search-dialects.

from tavily import TavilyClient

client = TavilyClient(api_key="<SpecStream key>", api_base_url="https://api.specstream.ai/tavily")
results = client.search("EU AI Act obligations for general purpose models", max_results=10)
pages = client.extract([r["url"] for r in results["results"][:3]])

The top published score

The Search API scores 97.9 F-score on all 1,000 SimpleQA Verified questions (978 correct) and 64.0 item F1 and 43.4 row F1 on all 200 WideSearch tasks. The Extract API scores 0.908 F1 on WCXB. Every score and its method is on the benchmarks page.

What a call keeps

Zero retention is set per API key: a search made with such a key keeps no request text and no answer after the reply, and the Extract API keeps no page. Another key setting sends searches out only from our server in Madrid, in the EU. Every call is in the account's request log with its price.

Jina Search and Reader

s.jina.ai searches and returns the top five results with their URLs and contents, and r.jina.ai reads one URL. Each search costs a fixed number of tokens, starting from 10,000, and every new key gets 10M free tokens. SpecStream answers both calls at https://api.specstream.ai/jina-search and https://api.specstream.ai/jina: see Jina Search compatible.

Exa

A search API with page contents in the reply: $4 per 1,000 requests on lite and $7 on auto, with up to 10 results. Its /contents endpoint is $1 per 1,000 pages per content type. SpecStream answers Exa's own POST /search and /contents at $1 per 1,000 searches and $0.80 per 1,000 pages: see Migrate from Exa.

Olostep

A web data API with an /answers endpoint that searches, reads and returns an answer with its sources, in 3 to 30 seconds by its own account. Plans start at $9 a month for 5,000 successful requests. SpecStream's search writes answer with its page and read time at $5 per 1,000, with no plan.

Firecrawl

Search at 2 credits per 10 results and scrape at 1 credit a page, with Standard at $99 a month for 100,000 credits. SpecStream answers Firecrawl's /v2/search and /v2/scrape on the same key: see Firecrawl migration.