The enterprise web search API buyer's guide for 2026

1. The price at your volume, in writing

List prices tell you little past a million searches a month. SpecStream prints its volume terms: 1,500,000 searches for $5,000 ($3.33 per 1,000) or 8,000,000 for $20,000 ($2.50 per 1,000), paid once and never expiring. On a signed order of 3 months or more, 1,000,000 searches a month cost $3,000 ($3 per 1,000) and 5,000,000 cost $10,000 ($2 per 1,000). A fast search draws a fifth of a default one from the pack. Exa's enterprise plan starts at a $12,000 annual commitment, Parallel's Scale annual plan is $10,000 charged upfront, and Linkup and Brave quote by contact. To pay by invoice or sign for committed volume, talk to a human.

2. Zero data retention as a setting on each key

Make a key with "zero_retention": true, or set it later with PATCH /v1/account/keys/{id}. A search made with that key keeps no request text and no answer after the reply: it is not in the history, its answer is not reused for another call, and its log lines carry neither. The charge keeps only the account, the key, the request ID, the price and the time, and the Extract API keeps none of the pages it returned. Linkup's documentation says ZDR is not enabled by default and is requested on demand, Exa adds it with an enterprise plan, and Parallel turns on Hybrid ZDR once an agreement is signed.

3. Paper your legal team can read before the first call

SpecStream SAS is a French company, and the service runs on servers in Madrid, Spain, with hosting and databases in the European Union. The data processing agreement is published, with each category of subprocessor, where it sits and the safeguard for any transfer, and account owners get an email at least 30 days before a subprocessor is added or replaced. A key made with "eu_egress_only": true sends every search out from Madrid only. Linkup's documentation lists infrastructure in the US, EU, Canada and APAC. No model is trained on your work; the security page covers the rest.

4. Accuracy you can rerun

SpecStream scores 97.9 F-score on all 1,000 official SimpleQA Verified questions (978 correct), and every raw answer is public on Hugging Face. It also scores 64.0 item F1 and 43.4 row F1 on all 200 official WideSearch tasks, and the Extract API scores 0.908 F1 on WCXB. Method and sources are on the benchmarks page. Before you sign with anyone, run 50 questions from the same set yourself. This script picks them with a fixed seed, so every vendor gets the same 50, and prints the expected answer beside ours, the p50 and p95 time and the spend:

import csv, io, json, os, random, urllib.parse, urllib.request

CSV = "https://huggingface.co/datasets/google/simpleqa-verified/resolve/main/simpleqa_verified.csv"
rows = list(csv.DictReader(io.StringIO(urllib.request.urlopen(CSV).read().decode())))
random.seed(42)
sample = random.sample(rows, 50)

took, spent = [], 0.0
for row in sample:
    url = "https://api.specstream.ai/v1/web-search?q=" + urllib.parse.quote(row["problem"])
    req = urllib.request.Request(url, headers={"Authorization": "Bearer " + os.environ["SPECSTREAM_API_KEY"]})
    reply = json.load(urllib.request.urlopen(req))
    took.append(reply["took_ms"])
    spent += float(reply["price_usd"])
    answer = (reply.get("answer") or {}).get("value")
    print(row["answer"], "|", answer)

took.sort()
print("p50", took[len(took) // 2], "ms  p95", took[int(len(took) * 0.95)], "ms  spent $", round(spent, 3))

Grade each line with the set's official grader prompt, then point the loop at each API on your list. On SpecStream the run costs about 25 cents of the $10 credit.

5. What happens at a limit, a retry or a failure

When an account has used its calls for the minute, the Search API answers 503 with Retry-After and charges nothing; the Python and TypeScript SDKs send it again by themselves when the wait is 10 seconds or less. An Idempotency-Key header makes a retry safe: a call is charged once per key. A call that returns no results costs nothing. Each key carries its own spending limit, and X-End-User-Cap-Cents caps what one of your own users can spend.

6. A request log your finance and security teams can read

Every Search API and Extract API call is in the account's request log, charged or not, with its key, status, latency_ms and price_usd, through GET /v1/account/requests. GET /v1/account/usage sums calls and spend by day, product or key, or by your own end user when you send X-End-User-Id, and returns CSV for finance. Single sign-on, SCIM and audit log export come with Enterprise, from $2,500 a month.

7. One supplier past the search

The same key and balance cover the Search API, the Extract API at $0.80 per 1,000 pages, where a page with no text is not charged, and Deep Search, which returns every match for a request as structured rows at 5¢, 15¢ or $1 a row under a maximum you approve before it runs. Source URL and read time on every value. A saved request reruns hourly to yearly, and each row comes back labeled changed, unchanged or no longer matching.

Tavily

A search API sold in credits at $0.008 each, pay as you go. A basic search uses 1 credit and an advanced search 2, so $8 or $16 per 1,000. Its Enterprise plan adds SLAs by contact. SpecStream charges $5 per 1,000 searches, or $1 on the fast tier, with zero retention as a setting on each key.

You.com

Its Web Search API costs $5 per 1,000 calls and its Contents API $1 per 1,000 pages, and its enterprise offer lists zero data retention and a DPA. SpecStream reads pages at $0.80 per 1,000, charges nothing for a page with no text, and publishes its DPA before you sign up.

SerpApi

A SERP API sold in monthly plans by search count, backed by uptime SLAs of up to 99.97%, with ZeroTrace Mode to keep no search parameters, queries or results. Each search returns a results page to parse. SpecStream returns the answer with its page and read time, typed rows and every result read, at $5 per 1,000 searches.