Budget Langfuse Cloud as a monthly plan fee plus units, where every trace, observation and score is one unit. Keep non-LLM spans out, choose the plan by how far back people must look at data, and set spend alerts before traffic grows. The checklist counts units the way Langfuse bills them.
When you need this
- You are moving from Hobby to a paid plan, or from a pilot to production traffic.
- Online LLM-as-a-judge evaluators are about to run on production traces.
- Langfuse shares an OpenTelemetry pipeline with Sentry, Datadog, or another tool.
Evaluation criteria
Count units
How many units does a month of traffic produce?
Look forRead a week of counts in Usage Management and project the monthly total. A run with 1 trace, 10 observations and 1 score produces 12 units.
Price the tiers
What does that volume cost on the plan you need?
Look forCore costs US$29 monthly and Pro costs US$199, each including 100,000 units. Additional units cost US$8 per 100,000 to 1 million, US$7 to 10 million, US$6.50 to 50 million, and US$6 above. Langfuse prices 5 million Pro units at US$551 monthly in its example. It bills the plan at the month's start and usage at the end.
Count evaluator output
Which features add units on their own?
Look forLangfuse bills every score as a unit, including scores from LLM judges, annotation queues and experiments. Each judge's model call also bills your own LLM provider. Point judges at the observations that matter, and leave the rest of the spans unscored.
Keep unrelated spans out
Is everything you send LLM work?
Look forFilter OpenTelemetry spans by instrumentation scope. One developer's single bot question produced 278 spans until a collector filter cut 79% of them.
Thin out high-volume paths
Do you need every trace?
Look forSet LANGFUSE_SAMPLE_RATE or an OpenTelemetry sampler. Sampling keeps or drops a whole trace with its observations and scores. Langfuse recommends keeping all traces where you can, because feedback and evals show a trace's value later.
Choose the plan by data access
How far back must people look?
Look forHobby gives 30 days of data access, Core 90 days, and Pro and Enterprise 3 years. Pro adds project retention policies with a 3-day minimum, and Langfuse deletes expired data nightly. Dataset items survive their source traces.
Set spend alerts
Who hears about spend before the invoice?
Look forOn Core and above, spend alerts email owners and admins when the expected invoice crosses a threshold. Langfuse checks every 60 to 90 minutes. Each alert fires once per billing cycle.
Plan for the add-ons
Which controls cost extra?
Look forPro's US$300 monthly Teams add-on includes enterprise SSO and enforcement, project RBAC, protected prompt labels, scheduled blob export, and a Slack channel. Enterprise costs US$2,499 monthly and adds audit logs and SCIM, plus SLAs. Custom volume rates and AWS Marketplace billing require an annual commitment.
Questions to ask vendors
- How many traces, observations, and scores does each service produce a month?
- Which evaluators run on production, and on which observations?
- Which non-LLM spans reach Langfuse today?
- How many days of history do engineers and reviewers need?
- Who receives spend alerts, and at what thresholds?
Red flags
- Every server span exported to Langfuse through a shared OpenTelemetry pipeline.
- LLM-as-a-judge on every observation with no filter.
- Comparing Langfuse's per-unit price with another tool's per-trace or per-GB price without converting to your own trace shape.
- No spend alert on a project that runs automated experiments.
Products in this piece
Sources
- Official documentation
- 7
- Public source
- 2
- Langfuse pricing
- Langfuse docs: Billable units
- Langfuse FAQ: How do I cut my Langfuse Cloud bill?
- Langfuse docs: Sampling
- Langfuse docs: Spend alerts
- Langfuse docs: Data retention
- Langfuse docs: LLM-as-a-judge
- Hugo Lesta on dev.to: Tracing A2A agents without burning your free tier (Aug 14, 2026)
- Hacker News comment preferring Langfuse over LangSmith (Aug 11, 2025)