Cloudflare's Web Search API vs Tavily: What Grounding an Agent Costs

If your agent needs live web grounding, Cloudflare’s new Web Search API does not change which search index answers your queries. It changes who bills you. The API, announced in early October 2026, routes search through AI Gateway and fronts three partner providers (Ceramic.ai, Exa, and Linkup) at their own list prices, “without any additional markup.” You are still picking an underlying index. What Cloudflare adds is a consolidated, usage-based billing path that sits next to your inference spend.

That framing matters for anyone currently standardizing on Tavily, Brave, Serper, or Exa directly, because it means the comparison is structural rather than a new entrant with a better index. One caveat up front: the only independent evaluation cited here, a 2026 arXiv preprint, does not cover Cloudflare’s offering at all and carries its own methodological confound. Everything quality-related about the Cloudflare path is currently vendor-said or unknown, and the product is days old. The cost and terms comparison below rests on documented pricing and published feature claims, each labeled by who said it.

What Cloudflare Actually Shipped: Partner Search Behind AI Gateway

Cloudflare’s launch post is explicit about the model: “We’re kicking off this launch with our partners from Ceramic.ai, Exa, and Linkup,” and “we also offer web search directly at list API pricing from our partners, without any additional markup.” Read that second sentence carefully. It is a pass-through statement, not a discount. There is no partner price in the announcement, so nothing in it supports a claim that routing through Cloudflare is cheaper than going direct.

What the gateway adds is plumbing: one credential surface, one billing relationship, and metering that folds into the same usage-based regime Cloudflare already runs for inference. On Cloudflare’s plans page, Workers AI bills paid usage at $0.011 per thousand neurons; the same page lists a free tier of 10,000 neurons per day. Search metered through AI Gateway becomes one more metered line item beside that, rather than a separate SaaS contract with a search vendor. For a team already running Workers and AI Gateway, that is a real operational simplification. For a team that is not, it is a reason to consider the gateway, not a reason to switch search providers.

Cloudflare also imposes compliance conditions on its partners. The launch post requires that “the crawler used by the web search provider must comply with Cloudflare’s publicly stated requirements for ‘Verified bots’” and that “web search responses must include a link to the location of crawled content.” The practical effect is that every response carries source links, which is useful for citation-building agents, and that the underlying crawlers play by Cloudflare’s published bot rules, which matters if your queries touch sites that gate crawlers through Cloudflare itself.

The Cost Model: What 1,000 Grounded Queries Costs, Vendor by Vendor

Here is the uncomfortable truth about building a cost table for this decision: almost nobody in this comparison publishes a per-query price in a form that survives scrutiny.

The one concrete anchor is Serper, and even that comes from a pricing breakdown published by a competing SERP API vendor rather than Serper’s own site, because serper.dev/pricing returned a 404 as of mid-July 2026 and pack pricing was visible only after signup. The publisher, Serpent API, states the conflict itself: “Serpent API is a competing SERP provider, so read this as an informed competitor’s buyer’s guide, not neutral journalism.” Its pack table is in turn drawn from a dated third-party breakdown (ColdIQ, seen July 17, 2026), which is worth knowing when you weigh a competitor’s numbers. Per that breakdown, Serper sells prepaid credit packs from $50 (50,000 credits) up to $3,750 (12.5 million credits). Those work out to roughly $1.00 down to $0.30 per 1,000 queries depending on pack size, per the same breakdown.

For Cloudflare’s partners (Ceramic.ai, Exa, Linkup), no per-query price appears in the launch post; get each partner’s list price before deciding. As of October 7, 2026, no Tavily or Brave per-query figure is documented in the pricing material gathered here. So the honest cost model looks like this:

That table is the correct shape of the decision even with its empty cells: it tells you what you still need to collect. The point of the exercise is not to crown a winner from incomplete pricing; it is to recognize that “Cloudflare adds no markup” and “Cloudflare is the cheapest option” are different claims, and only the first is documented.

Two Serper terms change its effective price enough to matter. One credit covers a query returning up to 10 results, but requesting 11 to 100 results costs 2 credits, per the same pricing breakdown. An agent that habitually pulls deep result sets doubles its per-query cost. And per the same analysis, paid credits expire six months after purchase, so the $0.30-per-1,000 tier only delivers that rate if you actually consume 12.5 million credits within the window. A team buying the largest pack to get the headline rate and using half of it has effectively paid $0.60 per 1,000, per the same breakdown. Cloudflare’s usage-based gateway billing has no equivalent cliff in what Cloudflare has published: you pay list price for what you use, when you use it. Whether that beats a well-utilized Serper pack at its cheapest documented rate depends on the partner list price you cannot yet see, and on how predictable your query volume is.

Latency and Index Quality: Vendor-Said Numbers vs the One Independent Study

The only latency comparison naming the incumbents among the sources cited here comes from Exa’s own homepage, which reports p50 figures of 178ms for Exa Instant, 260ms for “Tavily Ultra Fast,” and 559ms for Brave. Treat these as marketing measurements, not findings. Exa is selling the top row of that table. The numbers are plausible in shape (a dedicated low-latency tier beats general-purpose endpoints), but no independent replication appears in the sources cited here, and the comparison predates or ignores Cloudflare’s gateway path, which adds a network hop of unknown size. If latency drives your choice, the honest procedure is to measure p50 and p95 from your own region against your own query mix.

On result quality, the one independent study cited here is “Similar Accuracy, Unequal Evidence: Search APIs as Decision Surfaces for Tool-Using Agents,” an arXiv preprint that evaluated 100 questions from the SealQA-Hard subset across Brave, Tavily, and Firecrawl with a frozen GPT-5.4 agent, all sharing one page-fetch backend. Two things about it deserve attention.

First, its title is the finding: the providers it tested produced similar answer accuracy but differed in the evidence they surfaced. If that result holds, switching search providers alone does not fix grounding quality problems, which undercuts any comparison that treats index choice as the main lever. Second, the study’s own authors flag a real confound: “The main traces were collected sequentially on 2026-05-17 UTC: Brave 06:49–07:24, Tavily 07:32–08:11, and Firecrawl 08:32–09:02. Runs were not interleaved, so provider and collection-time effects are not fully separated.” Web indexes shift by the hour, so even its Brave-versus-Tavily differences carry a time-of-day effect. This is a single, unreplicated preprint, and it says nothing about Cloudflare, Exa, Linkup, or Ceramic.ai. Use it as a methodological template (frozen agent, fixed question set, shared fetch backend), not as a verdict on any provider you are considering.

Where Queries Go: Retention, Verified Bots, and Source Links

For agents handling customer data, where queries are processed can outweigh a tenth of a cent per query.

Exa advertises a Zero Data Retention policy: queries and results are never stored or trained on. That is a vendor policy statement, but it is at least a stated, checkable commitment, and it travels with Exa whether you reach it directly or through Cloudflare’s gateway. The launch post is explicit about the gateway layer as well: you can “collect logs and request data on web search calls”; “Requests show up in your normal AI Gateway observability logs”; and Cloudflare “will identify partners supporting Zero Data Retention (ZDR), so that you know that your data is not retained.” So search calls are logged at the gateway by default, and ZDR support will be labeled per partner. What the post leaves open is how long those logs are kept and whether Cloudflare offers contractual retention terms for them.

Cloudflare’s contribution on this axis is structural. Its Verified bot requirement constrains how partner crawlers behave, and the mandatory source links in responses give agent builders citation material by default rather than as an add-on. If your agent must justify its answers with links, that requirement does real work for you. If you need contractual guarantees about retention at the gateway layer, you will have to ask Cloudflare directly; the launch post commits to logging and ZDR labeling, not to retention terms.

Choosing: A Framework and a Verdict

Assembled from the documented evidence, the decision comes down to five questions, in roughly this order:

1. Are you already in Cloudflare’s billing orbit? If you run Workers and AI Gateway, gateway-metered search removes a vendor relationship for zero markup. If not, adopting the gateway for search alone means evaluating the whole platform for one API.
2. What does a thousand queries actually cost you, including terms? Serper’s documented $0.30–$1.00 range, per the competitor-published pricing breakdown, is the only hard anchor among the sources cited here, and it is conditional on consuming prepaid packs within six months and on the depth of results you request. Cloudflare’s number is whatever the partner charges; get the list price before assuming anything.
3. Does latency bind? Only if your agent searches in a user-facing loop. If it does, run your own measurements, because the only comparative latency figures in the sources cited here are Exa’s own marketing table.
4. What are your retention and citation requirements? Exa’s stated Zero Data Retention, Cloudflare’s commitment to identify ZDR-supporting partners, and the mandatory source links are the documented features here; everything else is unverified.
5. Can you tolerate being early? Cloudflare’s offering is days old, the one independent evaluation cited here does not cover it, no partner per-query price appears in the launch post, and its operational failure modes are unknown.

My recommendation follows from those. Treat Cloudflare’s Web Search API as a billing-and-consolidation decision, not a quality or price win. If you are standardizing today and cost is the binding constraint, Serper’s documented range makes it the anchor to beat, but model the 2-credit deep-query tier and the six-month expiry before trusting the $0.30 figure in the competitor-published breakdown, and confirm current pack pricing at signup given the 404 that breakdown documented. If you are already metered through AI Gateway, routing search through Cloudflare at partner list pricing is a reasonable default, provided you get the actual partner list prices in writing first.

Either way, gate standardization on a spot-check you run yourself: a fixed set of freshness-sensitive queries, issued interleaved across your finalist providers, scored for citation quality. The independent study cited here shows both that this is feasible and that even careful attempts carry confounds. The burden of verifying freshness and citation quality now sits with you, the agent builder. No vendor’s launch post was going to lift it.

Frequently Asked Questions

Does Cloudflare’s Web Search API charge a markup over partner list prices?

The API, announced in early October 2026, routes search through AI Gateway and fronts three partner providers (Ceramic.ai, Exa, and Linkup) at their own list prices, “without any additional markup.”

How long do Serper prepaid credits remain valid?

And per the same analysis, paid credits expire six months after purchase, so the $0.30-per-1,000 tier only delivers that rate if you actually consume 12.5 million credits within the window.

Does Exa store or train on search queries?

Exa advertises a Zero Data Retention policy: queries and results are never stored or trained on. That is a vendor policy statement, but it is at least a stated, checkable commitment, and it travels with Exa whether you reach it directly or through Cloudflare’s gateway.