What a web search API should cost in 2026

Kaer's Search API costs $0.005 per query — $5.00 per 1,000. No monthly minimum, no per-result surcharge, no separate line for page content. This post shows the price against every provider you are likely to be comparing us to, explains the mechanics that make headline prices misleading, and is honest about the one category where we are not the cheapest option.

Search is now the largest recurring line item in most agent workloads. A model call happens once per turn; a research agent might search fifteen times before it answers. So the per-query price stops being a rounding error somewhere around your first serious customer, and by the time you are running a million queries a month it is the difference between a $5,000 bill and a $10,000 one.

The trouble is that per-query prices in this market are quoted in ways that are hard to line up. One vendor charges per request, another per credit where a request can cost two credits, a third bundles search into a token bill. So before the comparison, one definition: everything below is normalised to what you pay for one thousand ordinary searches returning around ten results. Where a provider charges extra for going beyond that, the surcharge is called out in the notes rather than buried.

The price list, side by side

Prices read from each vendor's own published pricing page in August 2026. Costs are per 1,000 searches, in US dollars, excluding tokens.

Two things jump out of that table. The first is that the built-in search tools from the model providers are the most expensive way to buy search — you are paying a convenience premium for not having to hold a second API key. The second is that the cheap end of the market and the expensive end are not selling the same product, which is the next section.

The headline price is rarely the price you pay

Three mechanics account for most of the gap between a quoted price and an invoice.

Result-count surcharges. Exa's $7.00 covers ten results; every result past the tenth adds $1.00 per 1,000. Ask for twenty-five — a normal ask when you are feeding a reranker — and you are at $22.00 per 1,000, not $7.00. Serper's credits work the same way in miniature: one credit up to ten results, two credits for eleven to a hundred. Kaer returns up to twenty-five for the same $0.005, because the results were already in memory when the query ran.

Credit multipliers. “Per credit” pricing is only comparable to per-query pricing if a query is one credit. On Tavily, a basic search is one credit and an advanced search is two, so the $8.00 headline is $16.00 for the search depth that most retrieval-augmented generation pipelines actually want.

Plan minimums. Tavily's $5.00 rate is real, and it arrives with a $500-a-month commitment for 100,000 credits. SerpApi's $9.17 is real too, at $275 a month. If your volume is genuinely there, those are honest prices. If it is not, you are pre-buying searches you will not run — and the effective rate on the ones you do run is whatever the monthly fee divides down to.

One research task, priced end to end

Abstract per-thousand rates hide all of this, so here is a single concrete unit of work: one research query, twenty results, and the full text of five of those pages. That is roughly what one hop of a research agent costs.

Multiply by a million and the spread is $2,000 to $22,000 a month for the same unit of work. That spread, not the headline number, is what to put in the budget.

Where we are not the cheapest

A comparison post that concludes its author wins every category is an advert, so: if all you need is Google's result page, Serper is five times cheaper than us and you should use Serper. A dollar per thousand — thirty cents at volume — is a price we do not try to meet, because a SERP proxy is a fundamentally different thing to build. It fetches a search engine's results page on demand and parses it. There is no index behind it.

That difference cuts both ways, and it is worth being precise about which way it cuts for you:

• SERP proxies return metadata, not content. Title, URL, snippet. If your agent needs the page, you are buying a second product — a scraper or an extraction API — and paying its bill too. That is where the $0.002 quietly becomes $0.012.
• You inherit someone else's rate limits and blocks. A proxy's availability is a function of how the upstream engine feels about being proxied that week. Ours is a function of our own capacity.
• Latency is a round trip to a third party. A SERP proxy cannot be faster than the engine it proxies. Our web lane answers from our own index in about 80 ms.

Pick the SERP proxy when you want cheap links at volume. Pick a search API when the agent has to read what it finds.

Why the price is $5.00 and not $9.00

The price is structural, not promotional, and the reason is worth explaining because it tells you whether it will still be $5.00 next year.

Most search APIs are resellers. They hold a commercial agreement with an upstream engine, pay per query under it, and add a margin. Their floor price is somebody else's wholesale price, so it can move without warning when the contract is renegotiated.

Kaer runs its own crawler and its own index. When a query lands on the web lane, no money leaves the building. What we pay for is the hardware the index sits on, and that cost is amortised across every query rather than incurred per query — the more traffic the index serves, the less each query costs us. It is why undercutting the resold tier is achievable rather than a loss leader.

There is a fallback behind the index for the queries where our own coverage is thin. It fires on a minority of requests, and it is the reason the price carries a hedge rather than being lower still.

Three lanes, because not every query is the same shape

Charging one price for a 40-millisecond lookup and a three-second research run only works if you let the caller choose which one they are buying. So lane is a request parameter:

All three bill at $0.005 — the lane is a parameter on the request, not a plan you pick at signup, so a single key can run a cheap lookup and a deep read minutes apart. The rate limit on one API key is 60 queries per minute; if you need more than that, the limit is a conversation rather than a plan tier.

Calling it

One endpoint, bearer auth with a kaer_sk_ key carrying the search scope.

curl https://app.kaer.ai/api/v1/search \
  -H "Authorization: Bearer $KAER_API_KEY" \
  -H "content-type: application/json" \
  -d '{
    "query": "post-quantum key exchange in TLS 1.3",
    "lane": "deep",
    "num_results": 10
  }'

The response is { query, lane, results }. On the deep lane each result carries the extracted page text alongside the usual title, URL and snippet, plus the authority, freshness and likely-paywall signals the extractor produced.

Two responses worth handling explicitly. A 429 means you have passed 60 queries a minute on that key and carries a Retry-After. A 402 with search_api_requires_extra_usage means the account has not enabled extra usage — the Search API bills per query with no included allowance, so it needs that switch on before it will serve anything. It is a one-time setting, not a paywall you hit mid-run.

Or skip the integration entirely

If the thing you actually want is for Claude or ChatGPT to be able to search, you do not have to write any of the above. Kaer speaks the Model Context Protocol at /api/mcp: connect it once and your assistant gets the search index, your workflows, your mail and your calendar as tools, authenticated against your Kaer account and metered on the same bill.

It works with Claude Desktop, Claude Code, ChatGPT's custom connectors, Codex and Cursor — the full walkthrough is here. The button below creates the connection and hands you the config block for whichever one you use.

The short version

• $5.00 per 1,000 queries, flat. Up to 25 results, page extraction included on the deep lane, no monthly minimum.
• Half the price of the built-in web search tools from Anthropic and OpenAI, which are $10.00 per 1,000 plus tokens.
• Cheaper than Exa ($7.00 before surcharges) and Tavily ($8.00 pay-as-you-go), level with Brave.
• Not cheaper than a Google SERP proxy, and not trying to be — those sell links, not page content.
• The price holds because the index is ours, so a query costs us capacity rather than somebody else's per-query fee.

Frequently asked questions

How much does the Kaer Search API cost?

$0.005 per query, which is $5.00 per 1,000 queries. In Kaer credits that is half a credit per query. There is no monthly minimum, no per-result surcharge and no separate charge for page content on the deep lane.

Is there a free tier for the Kaer Search API?

No, and that is deliberate. The Search API bills per query with no included allowance, so it requires extra usage to be switched on for your account. A key on an account without extra usage would start returning 402 partway through a run with no way for you to anticipate it, which is worse than being told up front.

What is the cheapest web search API for AI agents?

On sticker price, Google SERP proxies such as Serper are cheapest at roughly $1.00 per 1,000 queries falling to about $0.30 at volume, but they return search-engine result snippets rather than page content. Among APIs built for agent retrieval — Kaer, Exa, Tavily, Brave and Perplexity — Kaer is the lowest at $5.00 per 1,000, tied with Brave's Search plan and undercutting Exa at $7.00 and Tavily at $8.00 pay-as-you-go.

How does the Kaer Search API compare to Anthropic's and OpenAI's built-in web search?

Both Anthropic's web search tool and OpenAI's Responses API web search are $10.00 per 1,000 searches plus the tokens the results consume. Kaer is $5.00 per 1,000 for the search itself. The built-in tools are simpler if you are already inside one model provider; a standalone API is cheaper and portable across providers.

How fast is the Kaer Search API?

It has three lanes. The web lane returns a direct index hit in roughly 80 ms at p50. The fast lane races several engines and returns in roughly 300 ms, or under a millisecond on a cache hit within the 30-minute window. The deep lane fetches and extracts the pages themselves and takes one to three seconds.

Can I use the Kaer Search API from Claude or ChatGPT?

Yes. Kaer exposes a Model Context Protocol server at /api/mcp that any MCP client can connect to with a Kaer API key — Claude Desktop, Claude Code, ChatGPT custom connectors, Codex and Cursor. The connection is authenticated against your Kaer account and metered on the same bill.