Datadog LLM Observability Review (2026)

The one thing to understand: spans, not traces

Nearly every mistake teams make with this product traces back to one billing detail.

Datadog bills per LLM span, not per trace. Each span inside a trace is priced independently. That sounds like a technicality. It is the difference between a $160 bill and a $4,000 one.

A simple completion request generates roughly 3 to 5 spans. An agentic workflow - tool calls, retrieval steps, reasoning chains - commonly generates 20 to 50. One published analysis puts AI agent workloads at 8 to 15 spans per request against 2 to 3 for a typical API endpoint.

So the free tier’s 40,000 LLM spans per month is not 40,000 requests. At 20 spans per agentic request it is 2,000 requests. At 50, it is 800. A modest-traffic agentic application can burn the entire monthly free allowance in a day or two.

If you take one thing from this page: count spans, not users, not requests. Instrument a representative request in a trial account and read the actual span count before you model anything.

Pricing, and where the sources disagree

Tier	Price	Included	Retention
Free	$0	40,000 LLM spans/mo	15 days
Pro	From $160/mo	100,000 LLM spans/mo	15 days
Enterprise	Custom	Volume commitments, regional	Custom

Two honest caveats on this table.

First, Datadog introduced new LLM Observability pricing effective 1 May 2026. Any comparison written before that date is stale, and there are a lot of those still ranking.

Second, third-party sources disagree on the current Pro rate. We found $160/month for 100,000 spans in some analyses and $240/month for the same allowance in others. We have published the lower figure with the conflict noted rather than picking one silently. Verify the current rate against Datadog’s own pricing page before budgeting - and note that Datadog states pricing varies by region, and that month-to-month and annual commitment plans are discounted differently, which is a plausible source of the discrepancy.

Retention add-ons are billed per 10,000 LLM spans on top of the base allowance.

The auto-activation trap

This one is worth flagging separately because of who it catches.

Multiple third-party analyses report that Datadog automatically activates the LLM Observability premium when it detects LLM spans in your account. The scenario is an existing Datadog APM customer whose product team ships an AI feature - spans start flowing, the premium activates, and charges appear that nobody explicitly authorised.

We could not confirm the precise activation behaviour against Datadog’s own documentation, so we are flagging this as reported-but-unverified rather than stating it as fact. Either way, the mitigation is the same: if you run Datadog and any team is adding LLM calls, talk to your account manager about LLM Observability billing before the code ships, not after the invoice.

What it is genuinely good at

The case for this product is correlation, and it is a real case.

Datadog has spent more than a decade building APM, and the LLM product inherits that. When an LLM span shows a latency spike, you can trace it down through the service, the host, the database query and the network - in the same product, on the same timeline. No AI-native observability tool does this, because none of them monitor your infrastructure.

The sampling controls are the other inherited strength. Head-based sampling decides at trace initiation; tail-based sampling keeps every error and slow request while sampling the fast successful ones. For LLM workloads tail-based is usually correct, because the traces worth keeping are the failures, and it is the single most effective lever on your bill. Most AI-native competitors have far cruder controls.

Built-in evaluations cover quality, safety, hallucination and PII detection, plus custom evaluations and human feedback via API. This is competent. It is not why you would buy the product, and it is thinner than Braintrust if regression testing is your priority.

The self-hosting answer is simply no

There is no self-hosted deployment. Not a gated tier, not an enterprise add-on - it does not exist.

This is a hard differentiator against Langfuse, Arize Phoenix and Opik, all of which run entirely inside your own infrastructure. Datadog offers regional deployments including the EU, which solves data residency in the jurisdictional sense, but your trace data still leaves your network and lands in Datadog. If you have a hard self-host requirement, this product is disqualified regardless of its other merits, and no amount of correlation value changes that.

Should you use it?

Use Datadog LLM Observability if you already run Datadog for infrastructure and APM, your LLM features live inside a larger system where correlation matters, and you have modelled your span volume honestly.

Don’t use it if you have no existing Datadog footprint, you need self-hosting, evals are your primary requirement, or you are running high-span-count agentic workloads on a tight budget.

Bottom line: this is an excellent LLM tab on a platform you either already pay for or should not buy for this reason alone. The correlation with infrastructure telemetry is genuinely unmatched. The per-span billing model is genuinely dangerous if you estimate it wrong, and agentic workloads are exactly the case where people estimate it wrong. Model spans first. If Datadog is already your monitoring platform the answer is probably yes; if it is not, look at Langfuse or Braintrust instead.

---

Pricing and features verified against public sources on 31 July 2026, following Datadog’s 1 May 2026 pricing change. This is a researched directory entry - we have not yet instrumented this platform with our reference application, so the span-count figures here are drawn from published analyses rather than our own measurement.