Datadog added LLM Observability to its platform for teams that already run their infrastructure monitoring there, and later made it available as a standalone product. The pricing is per span, in blocks of 10,000, with three billing modes that change the rate by more than 40%. This post sets out the tiers, the span, the three modes and the retention add-ons, then works three monthly bills.
Prices are as published on the Datadog pricing page in October 2026. Datadog quotes most prices at the annual rate, and regional pricing can differ, so treat these as list prices.
The tiers
| Free | Pro | |
|---|---|---|
| Price | $0 a month | From $160 a month annual, $200 month-to-month, $240 on-demand |
| Included | Up to 40,000 LLM spans a month | 100,000 LLM spans a month |
| Retention | 15 days | 15 days |
| Overage | Not available | Per 10,000 spans, see below |
Pro is standalone: Datadog states it can be purchased without any other Datadog subscription. The free tier needs no card.
What a span is
A span is one operation inside a trace: an LLM call, a tool call, an embedding, a retrieval, a workflow step. An agent run with six model calls and four tool calls is ten or eleven spans. A chatbot turn with one model call and a retrieval is two or three. Multiply your runs by spans per run to get the volume the bill is based on, the same arithmetic as Langfuse units, and unlike LangSmith's one-trace-per-run.
Overage and the three billing modes
| Billing mode | Base (100,000 spans) | Each additional 10,000 spans |
|---|---|---|
| Annual commitment | $160 a month | $3.50 |
| Month-to-month | $200 a month | $4.20 |
| On-demand | $240 a month | $5.00 |
The gap between annual and on-demand is 43% on the overage and 50% on the base, so the billing mode is the first thing to settle. A team trialling the product on-demand and projecting the bill at the annual rate will be a third out.
Retention add-ons
The default 15 days is short for debugging a customer complaint. Extended retention is priced per 10,000 spans per month:
| Retention | Per 10,000 spans a month |
|---|---|
| 30 days | $6.50 |
| 60 days | $8.00 |
| 90 days | $9.00 |
At 30 days the retention add-on is almost twice the annual overage rate, so keeping everything for a month roughly triples the per-span cost. The usual answer is to keep the default and rely on alerts to find problems inside the window.
Three worked months
Pro tier, default 15-day retention, first 100,000 spans in the base price:
| Monthly spans | Annual | Month-to-month | On-demand |
|---|---|---|---|
| 1,000,000 | $160 + 90 blocks at $3.50 = $475 | $200 + 90 at $4.20 = $578 | $240 + 90 at $5.00 = $690 |
| 10,000,000 | $160 + 990 at $3.50 = $3,625 | $200 + 990 at $4.20 = $4,358 | $240 + 990 at $5.00 = $5,190 |
| 50,000,000 | $160 + 4,990 at $3.50 = $17,625 | $200 + 4,990 at $4.20 = $21,158 | $240 + 4,990 at $5.00 = $25,190 |
Adding 30-day retention to the 10M-span annual month is another $6,500, taking it past $10,000. In run terms, at eleven spans per run, 10 million spans is about 900,000 runs: 30,000 a day, a mid-sized production agent.
Where the bill surprises people
- The quoted price is annual. On-demand is 43% more per span; check which mode you are on.
- Spans, not runs. A richly instrumented agent is ten or more spans per run, and auto-instrumentation records every one.
- Retention is the multiplier. Fifteen days is the default for a reason. Extended retention costs more than the ingestion it extends.
- OpenTelemetry attributes can trigger billing. Teams sending GenAI semantic-convention attributes through the OpenTelemetry pipeline have reported LLM Observability billing they did not expect. Filter before ingestion if you are not using the product.
What this buys you
If your infrastructure, logs and APM already live in Datadog, LLM Observability puts agent traces in the same console, with the same alerting and the same on-call integrations your team already uses, and that is worth a premium. The standalone option means a team that is not on Datadog can buy it alone, at the same per-span rates.
Where it fits less well is the question the span count never answers: was the output any good. Datadog's own documentation positions LLM Observability around traces, latency, errors and cost, and it does not score output quality against a rubric or page you when a scheduled agent stops running. Agent Ping is built around those: heartbeat and schedule monitoring, quality scoring on every run with no scorers to write, cost per customer and feature, on a flat monthly tier instead of a per-span meter. The Agent Ping vs Datadog comparison covers the detail, and the LLM token cost calculator estimates the model spend inside each run.