Search API pricing explained

Four incompatible pricing models

Flat per-request. One request, one price. Easiest to model and increasingly rare.

Credit-based. You buy credits and each endpoint consumes a different number. A basic search may be one credit, an advanced search two, and a deep research request anywhere from four to two hundred and fifty. The effective rate depends entirely on your call mix.

Per-request plus tokens. A search fee on top of token charges, so total cost depends on response length as well as call count.

Processor tiers. The same endpoint at several effort levels with prices spanning three orders of magnitude.

Where the headline rate stops being the rate

Surcharges are where a comparison quietly breaks. A vendor quoting $7 per 1,000 for search may add $1 per 1,000 for results beyond the tenth, $1 per 1,000 pages of content, and another $1 per 1,000 for summaries. A realistic RAG call touching all three costs substantially more than the headline.

The same applies to grounding APIs billed per search query where one prompt can trigger several queries — the per-prompt cost is a multiple of the per-query rate, and the multiple is not under your control.

How to normalise

Pick one operation and price it everywhere: a plain web search returning about ten results, no extraction, no summarisation, no research loop. That is the only unit that exists across every vendor in the category.

Then model your actual mix separately. If your agent fetches content for half its results, add the extraction cost per vendor. If it uses deep research, price that tier explicitly rather than assuming the base rate.

The costs that are not on the pricing page

Tokens. A result set that returns page bodies costs far more in model tokens than the API call itself. This regularly exceeds the search bill.

Extraction. If the search API returns links, something must fetch and clean the pages — a second vendor, or infrastructure you run.

Retries and follow-ups. Poor first results cause the agent to search again. Result quality is a cost line, not just a quality line.

Rate limits. A low ceiling that forces queuing has a real cost in latency and in the engineering to work around it.

Why index ownership shows up in the price

A reseller pays an upstream provider per query, so its floor is that cost plus a margin. A provider that owns its index pays for crawling, indexing and serving — largely fixed costs that amortise across query volume.

That is why own-index providers can be cheaper at volume, and why their unit economics improve as they grow while a reseller's do not.