How to spend less on data API calls: where the money really goes

Most of a data API bill comes from a handful of expensive calls and from asking for the same thing twice. In the last 30 days, 17,509 calls went through the FetchAPI gateway and cost $67.26 upstream. 64 of them (0.4%) cost $18.68, 28% of the bill, and 3,592 paid calls (about $11.56, 18% of spend) repeated a request that had already succeeded. Fix those two things before you shop for a cheaper provider.

Your own breakdown is on /account under "Where your money goes": spend by endpoint, repeat requests, and a cheaper endpoint for the same job where the catalogue has one.

Where the money went

The top endpoints by spend over the 30 days, from our own metering:

A $0.002 Google results page is not where the money goes. A $0.29 AI-mentions call is, because it is priced per row.

1. Ask for the rows you will read, not the maximum

DataForSEO's LLM Mentions search charges $0.10 per request plus $0.001 per row returned. A call with limit: 1000 costs up to $1.10; the default 100 rows cost $0.20. In our 30 days, 48 calls with limit: 1000 cost $15.88, 24% of all spend.

We made the call with limit: 10 for rb2b.com: it cost $0.11 and told us there are 112 mentions in total (total_count), so we knew exactly how many rows a full pull would buy before buying them.

The same pattern holds wherever a price has a per-item part: DataForSEO Labs Ranked Keywords is $0.012 per task plus $0.00012 per keyword, so a 1,000-keyword answer costs $0.132 and a 100-keyword answer about $0.024.

2. If you only need the count, call the summary endpoint

Many APIs have a cheap "how many" endpoint next to the expensive "list them all" one. For rb2b.com, llm_mentions/aggregated_metrics returned the mention totals by platform and location for $0.101: one row, no list. If the question is "is this brand mentioned in AI answers, and how much", that is the call. Pull the rows only for the targets that clear your threshold.

3. Cache what does not change between calls

A company profile, an app record or last month's search volume does not change in an hour. Of the 30 days' paid calls, 1,330 distinct requests were sent more than once, 3,592 repeats in all: the second and later copies cost about $11.56 and returned what the first one had. Store each answer with the time you fetched it and reuse it for as long as the data can't have moved (a day for profiles, a month for monthly metrics).

# one call, saved; re-read the file instead of calling again
curl -s "https://fetchapi.co/api/gw/dataforseo/ai_optimization/llm_mentions/aggregated_metrics/live" \
  -H "x-api-key: $FETCHAPI_KEY" -H "content-type: application/json" \
  -d '[{"target":[{"domain":"rb2b.com"}],"location_code":2840,"language_code":"en","platform":"google"}]' \
  > cache/rb2b.com-llm-mentions-$(date +%F).json

4. Use the queue when you don't need the answer this second

DataForSEO prices the same Google results page three ways: $0.0006 in the standard queue (task_post, then task_get when it's ready), $0.0012 with priority, and $0.002 live. A nightly rank tracker has no reason to pay the live price: the queue is 3.3x cheaper for the identical page.

5. Pick the cheapest endpoint for the job, not the first one you find

The same job often has several routes in the catalog at very different prices. Listing a TikTok user's latest posts costs about $0.00002 a call through tiktok-scraper7 and about $0.002 through tiktok-api6: a 100x difference for the same list. We wrote up why in Why the same TikTok data costs 100x more from one API than another. On /account, any endpoint you use that has a cheaper same-job route is flagged with the price.

6. Watch the calls whose price isn't fixed

Some prices depend on what you send. DataForSEO's LLM Responses (ChatGPT, Claude, Gemini answers) charge a task fee plus the model's own token charge, so a long prompt or a long answer costs more. Our 95 ChatGPT-response calls averaged $0.065 each. Keep prompts short and cap the output length when the API lets you.

7. Don't pay for failures, and don't retry blindly

FetchAPI does not charge for a failed call (terms), and the gateway reads the body too, so a provider answer that says HTTP 200 but carries an error is treated as a failure. Retrying a request that failed because the input was wrong just fails again; retry only on timeouts and 5xx, with a back-off. The hidden costs post covers retries and failover in detail.

A checklist before you run a job at scale

1. Run it on 10 inputs and read the upstream cost per call (every gateway answer carries it in meta.cost_usd).
2. Set limit / depth to what you will read.
3. Use the summary endpoint if you only need counts.
4. Cache by request; reuse within the data's natural refresh period.
5. Use queued modes for anything that can wait.
6. Check /account after the first day: the "Where your money goes" table shows the top endpoints and any cheaper same-job route.

Get a key on /account, browse prices per call in the catalog, or connect the MCP server and let your agent call get_usage to see its own spend.