Building a tool router means I pay for the same web search from six different vendors, which gives me an unusual vantage point: the bill. Search APIs are the most commoditized category in the agent stack, and also the one with the widest price spread. The cheapest option on our catalog costs $1.20 per thousand queries routed. The most expensive costs $8.40. That is a 7x difference for a tool call most agents fire dozens of times per session.
The spread is not irrational, though. These six products sit in two distinct tiers, and knowing which tier your workload needs is worth more than any individual pick.
The SERP tier: cheap, fast, navigational
1. Serper
Serper is a Google SERP API and it is the price floor of the category: $1.20 per 1k queries through our router, $1.00 direct. If your agent’s searches look like “cloudflare durable objects pricing” or “openai status page”, meaning you want the same links a human would get from Google, Serper is close to unbeatable on cost. It is where our default cheapest-first routing sends search traffic, and its quality score in our catalog (72) reflects exactly what it is: Google’s results, nothing more, nothing less.
2. Brave Search
Brave runs its own independent index, which matters more than it sounds. When Google-derived APIs wobble or get rate-limited, Brave keeps answering, and an independent index is the failover you actually want (two Google-SERP vendors can fail for the same upstream reason). At $6.00 per 1k routed it is mid-pack on price with solid freshness support. We score it 80.
3. You.com
You.com’s search API returns snippets designed to be fed to language models, with up to 100 results per call, and at $6.00 per 1k routed it is priced identically to Brave in our catalog. The LLM-shaped payloads are the draw: if you are stuffing results straight into a context window, you spend fewer tokens cleaning them up. We score it 78.
The neural tier: semantic, research-grade
4. Exa
Exa is the quality ceiling of our search catalog (score 90) and the price ceiling too, at $8.40 per 1k routed. It is a neural search engine: you can ask it conceptual questions (“startups building agent sandboxing”) and get documents that match the meaning, not the keywords. It also returns text content with results, which can save you a scrape call entirely. For research agents, that two-in-one often makes Exa cheaper in practice than its sticker price suggests.
5. Perplexity
Perplexity’s Search API arrived with their name recognition and slots into the neural tier at $6.00 per 1k routed with a score of 86. Results come back with dates and clean snippets, and freshness filtering works well. It is the strongest “default quality” pick if you want one provider that handles both navigational and semantic queries acceptably.
6. Firecrawl Search
Firecrawl is better known for scraping, and their search endpoint’s superpower is the same as Exa’s: it can return full page content with the results. At $2.40 per 1k routed it is the budget pick of the content-included options. We score it 74; the results skew utilitarian, but “search plus content in one call” at near-Serper prices is a real workload fit.
How I would actually route this
The honest answer is that no single pick is right, because search quality is query-dependent. Navigational lookups do not need Exa. Semantic research queries waste money on Serper retries. This is exactly why route.tools exists:
- Default: cheapest-first with failover. Serper wins, Brave catches its failures on an independent index.
- Research agents:
routing: { "sort": "quality" }on the hard queries, cheapest everywhere else. The routing docs show how to set a per-category default once. - Skeptics: pass
include_raw: trueand inspect what each provider actually returned before you commit to one.
Full pricing for all six, side by side and generated from the same open-source catalog our router reads, lives on the search API comparison page. And if you would rather not pick at all, one key routes across all of them with automatic failover.