What is Datadog LLM Observability?
Datadog LLM Observability, which Datadog now sells as Agent Observability, records every step of an LLM app or agent as a trace: prompts, retrieval, tool calls, latency, tokens and cost. On top of the traces it runs managed and custom LLM-as-a-judge evaluations, scans for prompt injection and sensitive data, groups traffic into topics, and offers datasets, experiments, a prompt playground and annotation queues for human review.
Apps are instrumented with the Python or Node.js SDK, the Java tracer, OpenTelemetry or an HTTP API, and the SDK captures OpenAI, Anthropic, Bedrock, LangChain, CrewAI and LiteLLM calls without code changes. It can be bought on its own, but its distinctive feature is linking an agent's trace to the services, infrastructure and real-user sessions already in Datadog.
Billing counts only LLM spans, meaning individual calls to a model provider; tool, retrieval and workflow steps are free. The Free tier covers up to 40,000 LLM spans a month with 15-day retention. Pro starts at $160 a month on an annual commitment ($200 month-to-month, $240 on demand) for 100,000 spans, then $3.50 to $5 per 10,000.
How much does Datadog LLM Observability cost?
| Plan | Price | What you get |
|---|---|---|
| Free | Free | Up to 40K LLM spans/month, 15-day retention, evaluations included |
| Pro | From $160/month annual | $200 month-to-month or $240 on demand for 100K LLM spans; then $3.50 to $5 per 10K LLM spans |
Free up to 40K LLM spans/month with 15-day retention. Pro from $160/month annual, $200 month-to-month or $240 on demand, for 100K LLM spans; then $3.50 (annual), $4.20 (month-to-month) or $5 (on demand) per 10K LLM spans. Retention add-ons listed at $6.50, $8 and $9 per 10K spans a month for 30, 60 and 90 days. Only LLM spans are billed.
Is Datadog LLM Observability worth it?
When a customer says your chatbot gave a wrong answer and was slow about it, the cause might be the prompt, the model, a slow database or a busy server. Most AI tracing tools only see the first two.
Datadog's tool records each step your agent takes and puts it on the same timeline as the rest of your systems, so the slow answer and the overloaded database show up side by side. If your engineers already use Datadog for servers and apps, that link comes without any extra setup, which none of the specialist tracing tools we cover can match.
The way it charges is easy to forecast. You pay only for calls to the model, not for every tool call or search step around them. An agent that makes five model calls per conversation uses 500,000 spans for 100,000 conversations. On an annual Pro plan, a million spans a month costs $160 plus 90 blocks of 10,000 at $3.50, about $475; the same month on demand is about $690. Evaluations are included in the span price, so you are not charged twice for scoring traffic.
The catch: what you save on Datadog you can spend elsewhere. The LLM-judge evaluations run on your own OpenAI, Anthropic, Bedrock or Vertex account, so each judged trace is a separate model bill; the docs suggest sampling, for example 10%, to keep it down. Traces are kept only 15 days unless you buy longer retention, listed at $6.50 to $9 per 10,000 spans a month for 30 to 90 days, and you have to ask support to change it. The best prices need a yearly commitment. Datadog lists its certifications in a Trust Center you have to request access to, and the product is not available on its US government sites.
Our take: this is the tracing tool you choose for your operations team as much as your AI team. A startup that does not already pay Datadog gets more from Langfuse or Braintrust; a company that does is better off keeping agent traces next to everything else.
Where it earns its place
An SRE team at a 300-person e-commerce company already on Datadog APM
Work out whether slow chatbot answers come from the model or our own services
Turn on the SDK in the chatbot service and its traces share a trace ID with existing APM spans, so a slow answer opens straight onto the database or API call behind it. Pro at $160 a month annual covers the first 100,000 model calls.
A two-person team building an internal HR assistant
See what employees ask and catch bad answers before launch
The Free tier's 40,000 LLM spans a month covers a pilot; topic grouping shows what people ask, and a sampled LLM judge on your own OpenAI key flags weak answers. Traces disappear after 15 days, so save examples to a dataset.
A head of platform at a fintech handling customer messages
Keep account numbers and personal data out of stored traces
Sensitive Data Scanner is included, with 1 GB of scanning per 10,000 spans, and SDK span processors can redact or drop data before it leaves the app. Access to each AI app's traces can be limited by team.
Before you commit
- Setup
- Install the Python or Node.js SDK (or use the Java tracer), set a Datadog API key and app name; OpenAI, Anthropic, Bedrock, LangChain, CrewAI and LiteLLM calls are traced without code changes. Datadog says most teams see traces within minutes.
- Integrations
- OpenAI, Anthropic, Gemini, Vertex AI and Bedrock; LangChain, CrewAI, Pydantic AI, LiteLLM and Strands Agents; OpenTelemetry GenAI conventions; HTTP API; export API for outside evaluators; NeMo; CLI and MCP access.
- Data and privacy
- Traces kept 15 days by default, or 30, 60 or 90 with an add-on; datasets and prompts kept 3 years; metrics 15 months. Sensitive Data Scanner and per-app access controls included. Certifications are in a Trust Center behind an access request, and the docs do not say whether trace data is used for training.
- Cost at scale
- One million LLM spans a month is about $475 on an annual Pro plan, $578 month-to-month or $690 on demand, before longer retention and the judge model's own token bill.
- Limits
- The Free tier cannot buy retention add-ons; experiment results are kept 15 days on free and on-demand plans; the product is not offered on app.ddog-gov.com or us2.ddog-gov.com.
What’s good
- Agent traces linked to Datadog APM, infrastructure and real-user sessions
- Bills only on model calls; tool and retrieval steps are free
- Free tier of 40,000 LLM spans a month
- Evaluations, experiments, datasets and annotation in every tier
- Sensitive data scanning included
What’s not
- LLM-judge evaluations run on your own provider account and bill
- 15-day trace retention by default; longer costs extra
- Lowest prices need an annual commitment
- Retention changes go through support, not the UI
- Not available on Datadog's US government sites
Best for
Engineering teams already running Datadog who want agent traces next to their service and infrastructure monitoring.
Datadog LLM Observability alternatives
Datadog LLM Observability: common questions
- Is Datadog LLM Observability worth it?
- If your company already lives in Datadog, this is the sensible place to watch your agents, and its per-model-call billing is fairer than it first looks. We rate it 4.0 out of 5.
- How much does Datadog LLM Observability cost?
- Free up to 40K LLM spans/month with 15-day retention. Pro from $160/month annual, $200 month-to-month or $240 on demand, for 100K LLM spans; then $3.50 (annual), $4.20 (month-to-month) or $5 (on demand) per 10K LLM spans. Retention add-ons listed at $6.50, $8 and $9 per 10K spans a month for 30, 60 and 90 days. Only LLM spans are billed.
- What are the downsides of Datadog LLM Observability?
- LLM-judge evaluations run on your own provider account and bill; 15-day trace retention by default; longer costs extra; lowest prices need an annual commitment; retention changes go through support, not the UI; not available on Datadog's US government sites.
- Who is Datadog LLM Observability best for?
- Engineering teams already running Datadog who want agent traces next to their service and infrastructure monitoring.
- How much does Datadog LLM Observability cost for a team or in production?
- One million LLM spans a month is about $475 on an annual Pro plan, $578 month-to-month or $690 on demand, before longer retention and the judge model's own token bill.
- Is Datadog LLM Observability safe for company data?
- Traces kept 15 days by default, or 30, 60 or 90 with an add-on; datasets and prompts kept 3 years; metrics 15 months. Sensitive Data Scanner and per-app access controls included. Certifications are in a Trust Center behind an access request, and the docs do not say whether trace data is used for training.
- What are the best Datadog LLM Observability alternatives?
- The closest alternatives we have reviewed are LangSmith, Braintrust and Arize AX.
Answers checked Nov 2, 2026.
Make Datadog LLM Observability? Show this rating on your site
