If your agent has to answer anything newer than its training data, it needs web search, and since Microsoft shut down the Bing Search APIs on August 11, 2025, picking a replacement is a real decision. The market is crowded and flush. In May 2026 Exa raised $250 million and was valued at $2.2 billion; in April, Parallel Web Systems took a $100 million Series B at a $2 billion valuation; and in February, Nebius agreed to acquire Tavily.
I priced one workload across all of them. The same 100,000 searches a month can cost you $100 or $1,300. I wrote up the full comparison, with every vendor section and the source for each price, on DevToolLab. This is the short version.
The number that decides your bill
Forget the headline price per 1,000 queries for a second. The first question is what comes back.
A Google-results API like Serper hands you titles, URLs and snippets. If the model needs to read the page, you make another call per URL to fetch and clean it, and that second hop costs money too. Exa, Tavily, Parallel and Perplexity return extracted page text or relevant excerpts in the same response. Brave covers both cases: call its standard Web Search endpoint for links plus snippets, or its LLM Context endpoint for page content that is already extracted, and both cost $5 per 1,000.
The second question is depth. Most vendors sell a quick mode for tool calls inside a chat loop and a heavier one that does more retrieval. At Parallel and Perplexity the heavy mode is 5x the price; at Tavily it's 2x.
The workload
A support-research agent for a US SaaS app, running in us-east-1, making 100,000 searches a month at 10 results each. When an API only returns snippets, the agent pulls full text for the top 3 hits (300,000 pages), at $1 per 1,000 pages, which is what both Parallel Extract and Exa Contents charge. Free monthly credits come off the top. Prices are from each vendor's own pricing page as of September 30, 2026.
# Monthly cost of 100,000 agent searches, list prices as of Sept 30, 2026.
SEARCHES = 100_000
FETCH = 300 * 1 # 300k pages at $1 per 1k, only for APIs that return snippets
tavily = lambda credits: 500 + max(0, credits - 100_000) * 0.008 # Growth plan
costs = {
"Parallel fast": 1 * SEARCHES / 1000,
"Perplexity fast": 1 * SEARCHES / 1000,
"Exa instant": 4 * SEARCHES / 1000 - 10,
"Serper + fetch": 1 * SEARCHES / 1000 + FETCH,
"Brave LLM Context": 5 * SEARCHES / 1000 - 5,
"Tavily basic": tavily(SEARCHES),
"Exa auto": 7 * SEARCHES / 1000 - 10,
"Tavily advanced": tavily(SEARCHES * 2),
}
for name, usd in sorted(costs.items(), key=lambda kv: kv[1]):
print(f"{name:<18} {'$' + format(usd, ',.0f'):>7}")
Parallel fast $100
Perplexity fast $100
Exa instant $390
Serper + fetch $400
Brave LLM Context $495
Tavily basic $500
Exa auto $690
Tavily advanced $1,300
Swap in your own volume before trusting any of these rows. I didn't score result quality; that depends too much on your queries to generalize.
Exa
Exa runs its own index and bills by how hard each query works: instant is $4 per 1,000, fast and auto are $7, and Deep Search runs $12 to $15. Page contents are included for up to 10 results, and each extra result is $1 per 1,000. The free tier is $10 a month (about 2,500 instant searches, no card), and volume discounts mean talking to sales.
Tavily
Tavily bills in credits: a basic search is 1 credit, advanced is 2. Pay-as-you-go is $0.008 a credit, and plans go from $30 for 4,000 credits up to $500 for 100,000. You get 1,000 credits a month free with no card, and failed URL extractions aren't charged. The site now brands it "Tavily by Nebius", so keep an eye on the roadmap.
Brave Search API
Brave's results come from its own crawler, not Google or Bing, and its API page puts the index at more than 30 billion pages. Search and LLM Context both cost $5 per 1,000 (50 queries per second on search), and a separate Answers endpoint is $4 per 1,000 plus $5 per million tokens. You get $5 of credit each month, but even the free plan wants a credit card.
Parallel
Parallel, the company former Twitter CEO Parag Agrawal founded, charges $1 per 1,000 in turbo and fast modes and $5 in basic and advanced, with 10 results and excerpts included. The excerpts are compressed passages, not whole pages. Extract is $1 per 1,000 URLs if you need the full text. Its pricing page describes the free allowance two different ways, so check your dashboard.
Perplexity Search API
Perplexity sells raw search results, ranked and with page content pulled out, for $5 per 1,000 calls, or $1 per 1,000 if you switch on Fast Search. One call can return as many as 20 results. search_context_size sets how much text comes back per result. There's no free monthly allowance listed, and if you find an old Sonar tutorial, note that the docs now point Sonar users to the Agent API.
Serper
Serper returns Google's own results as JSON. Credits are prepaid and last 6 months: $50 gets you 50,000 queries, and the largest pack works out to $0.30 per 1,000, the cheapest per-query rate here. The catch is no page content, which adds $300 of fetching in this workload, plus some legal weather: Google sued SerpApi, another Google-results vendor, in December 2025, and SerpApi moved to dismiss in February 2026.
If you do fetch pages yourself, DevToolLab's HTML to Markdown converter is a quick way to check how a page will look once it's cleaned up for a prompt.
SearXNG, if you'd rather self-host
SearXNG is the open-source route: an AGPL-3.0 metasearch engine (37,779 GitHub stars on September 30, 2026) that queries other engines for you. JSON output is off by default, so add json under search.formats in settings.yml. It has no index of its own, and its default config suspends upstream engines that start answering with CAPTCHAs or rate-limit errors, which heavy traffic from one IP will trigger. Good for dev and low volume.
How I'd choose
- Tight budget, agent loop: Parallel or Perplexity in fast mode, $100 for 100,000 searches with content.
- One API from quick lookups to deep research: Exa.
- Free tier, no card: Exa or Tavily.
- You need Google's ranking specifically: Serper plus an extract step, after reading up on the SerpApi case.
- You want an index that isn't Google or Bing: Brave.
- Zero budget and light traffic: a SearXNG instance on a cheap VM.
Whichever you shortlist, run 50 of your real queries through the top two before committing. And remember that search results land in your prompt: the LLM Token Cost Calculator shows what 10 results of context add to each turn. The full breakdown has each vendor's details and the sources behind every number.
Top comments (0)