Why your Datadog bill keeps climbing, and the line to cut first

Datadog rarely gets expensive in one jump. It creeps, because the bill is not one price but several separate meters that each grow on their own: hosts for infrastructure, hosts again for APM, and logs billed twice, once to ingest and once to index. Add a noisy service or a chatty log stream and one meter runs away while you were watching another.

The meters that make up the bill

• Infrastructure: about 15 dollars per host a month on the annual Pro plan. Predictable, scales with your fleet.
• APM: about 31 dollars per host a month, usually on a subset of those hosts. Tracing is the pricier per-host line.
• Log ingestion: about 0.10 dollars per GB sent. Cheap per unit, but volume adds up fast.
• Log indexing: about 1.70 dollars per million events to keep them searchable at 15-day retention. This is the line that surprises people. (List prices, checked June 2026, they move.)

Logs are where the bill runs away

Hosts are easy to forecast; logs are not. Because ingestion and indexing are separate charges, a verbose service costs you on the way in and again to keep it queryable. The fix is rarely fewer hosts, it is ingesting broadly but indexing selectively: keep only the log events you actually search at full retention and let the rest stay cheap. On most bills, moving indexing down moves the total more than removing a server would.

Flex Logs: the lever that is not just indexing less

The old advice was binary: index a log at full price or drop it. Datadog Flex Logs adds a third option, a cheaper storage tier for logs you must keep but rarely search. Standard indexing keeps logs searchable at about 1.70 dollars per million events at 15-day retention; Flex Storage keeps them queryable at about 0.05 dollars per million events, roughly a 34 times lower rate on the index line, with retention up to 15 months. The trade is that Flex charges a separate retrieval cost when you actually query it, so it suits audit trails, compliance logs and the long-tail debugging you touch once a quarter, while the logs behind live dashboards and alerts stay in standard indexing. On a bill where indexing is the runaway line, moving rarely-read logs to Flex instead of deleting them is often the largest single cut, and it keeps the data you were about to throw away. (List prices, checked July 2026, they move.)

The custom-metrics trap, and the lever that pulls it back

Each Infrastructure Pro host includes about 100 custom metrics, and overage is billed by usage, with Datadog not publishing one fixed overage rate. The count is easier to blow past than it looks, because of how Datadog counts one. A custom metric is not one line on a graph, it is one unique combination of a metric name and its tag values, what Datadog bills as a single timeseries. So request.latency tagged only by endpoint and status is a handful of metrics, but the same metric also tagged by user id, request id or container id becomes one metric per user, per request, per container. A single high-cardinality tag can turn one custom metric into thousands, which is why this line runs away without adding a host, and why a bill that jumped with no new servers is usually a tag someone added, not traffic.

The lever is the same ingest-broadly, keep-selectively move you pull on logs and spans, and here it has a name: Metrics without Limits. You keep sending Datadog every tag, but you configure which tags stay queryable, and only that queryable set counts toward your billable custom metrics. Your Usage page then shows two volumes, ingested custom metrics (everything you send) and indexed custom metrics (the queryable slice you pay the cardinality price on), so dropping a runaway tag like request id from the queryable set pulls the billed count down while the data keeps flowing in. Audit which tags you genuinely query before paying to keep all of them addressable.

APM is not just per host: spans are metered too

The roughly 31 dollars per host is only the entry charge on APM. Like logs, tracing carries its own volume meters stacked on top of the host line, and they are why an APM bill can climb while your host count sits still. Every APM host includes 150 GB of ingested spans and 1 million indexed spans a month; past those allotments Datadog bills about 0.10 dollars per additional GB ingested and about 1.70 dollars per additional million indexed spans, the same ingest-then-index split you pay on logs. A chatty service, a traffic spike or a new high-throughput endpoint can burn through the included spans without adding a single host, so the trace line grows on its own. The lever is the retention filter: it decides which spans get indexed and therefore billed, so index the traces you actually query (errors, slow requests, the routes you debug) and let the rest be ingested but not indexed. As with logs, tuning what you index usually moves the APM total more than removing a host would. (List rates, checked July 2026, they move.)

How hosts are counted: the high-water mark

Datadog does not bill on your average host count, or even on the number running right now. It bills on a high-water mark. It meters your host count every hour, drops the top 1 percent of hours in the month (about 7 hours out of roughly 720), and bills the whole month at the highest count in the remaining 99 percent. That top-1-percent exclusion forgives a handful of isolated peak hours, but it does not forgive a sustained spike: a migration or a load event that doubles your fleet for a couple of days sets the bill for the entire month, long after those extra hosts are gone. This is why a bill can climb even when your steady-state host count did not.

Commitment versus on-demand

The roughly 15 dollars per host is the annual-commitment rate. Month to month, the same Infrastructure Pro host is about 18 dollars, and any usage above your committed host count is billed at that higher on-demand rate, which across Datadog products tends to run 20 to 50 percent over the committed price. The commitment cuts both ways: under-commit and you pay a premium on every host you overflow into on-demand; over-commit and you pay all year for hosts you never ran. Because the bill keys off your high-water mark and not your average, forecast the commitment against your 99th-percentile host count. (List prices, checked June 2026, they move.)

The newest meter: watching an AI feature runs on its own line

If you added Datadog to watch an LLM or agent feature, that cost does not ride on the host or APM meters above. LLM Observability, which Datadog now lists under Agent Observability, is a separate SKU with its own unit: the LLM span. Datadog defines that span narrowly, a single call to an LLM provider such as OpenAI or Anthropic, and the tool, workflow, agent, embedding and retrieval spans around it are free. So an agent that makes three provider calls to answer one request counts as three billed spans, not the dozen orchestration steps around them. That is generous on complexity and unforgiving on volume: the line tracks how often your product calls a model, not how many hosts you run.

The list rate (annual commitment, checked July 2026, it moves): the first 100,000 LLM spans a month are 160 dollars, and each additional 10,000 spans is 3.50 dollars. Month to month those are 200 dollars and 4.20 dollars; on demand, 240 dollars and 5 dollars. Keeping spans queryable longer than the default adds roughly 1.50 dollars per 10,000 for 30-day retention, 3 dollars for 60 and 4 dollars for 90. Worked through, an app making 500,000 provider calls a month sits at 160 dollars for the first 100,000 plus 40 blocks of 10,000 at 3.50 dollars, about 300 dollars a month on the annual rate, and it climbs with usage the same way the log line does. Because only provider calls are billed and the orchestration around them is not, the way to hold this line down is fewer or cheaper model calls (cache repeat prompts, route easy steps to a smaller model, batch where you can), not trimming the workflow. The estimator on this page covers the host, APM and log meters; this AI line sits on top of that total as its own SKU, and it will not show up anywhere in your host count.

See your own number

The honest way to plan is to add the meters up with your real fleet and log volume. Put your host count, APM hosts and monthly log GB in and the tool estimates the monthly bill, shows which line dominates, and lets you edit every rate, so you can see the effect of indexing less before you renew. List pricing is only a baseline, your on-demand or negotiated rate will differ.