Langfuse is the open-source LLM observability platform, and its hosted version, Langfuse Cloud, is priced by the unit. The pricing is clearly published and has no per-seat component, which is why it is popular with the teams that find LangSmith's seats annoying. The catch is what a unit is, and how many of them one agent run produces. This post explains the plans, the unit, the graduated overage, and works three monthly bills.
Prices are as published on the Langfuse pricing page in October 2026.
The four plans
| Plan | Price | Included units | Data access | Users |
|---|---|---|---|---|
| Hobby | $0 | 50,000 a month | 30 days | 2 |
| Core | $29 a month | 100,000 a month | 90 days | Unlimited |
| Pro | $199 a month | 100,000 a month | 3 years | Unlimited |
| Enterprise | $2,499 a month | 100,000 a month | 3 years | Unlimited |
Core, Pro and Enterprise include the same 100,000 units; the difference is data access, support and features such as SSO, data retention management and the Pro Teams add-on at $300 a month. Hobby has no overage, so it stops at 50,000 units. Self-hosting is free: Langfuse is open source and ships Docker Compose and Kubernetes templates.
What a unit is
A unit is one of three things: a trace, an observation or a score. Observations are the spans, generations and events inside a trace. This is the part that decides the bill.
Take an agent run that calls the model six times and runs four tools. Instrumented the usual way, that is one trace plus six generation observations plus four span observations: eleven units. Add an LLM-as-judge score and a user feedback score: thirteen. A chatbot turn with one model call is two or three units. A deep research agent with forty steps is forty-plus.
So your monthly unit volume is runs multiplied by units per run, and the second number is between two and fifty depending on the agent. A million runs a month of a typical tool-using agent is ten to fifteen million units.
Graduated overage
Overage above the included 100,000 units is graduated, with each band priced separately:
| Units in the month | Rate per 100,000 |
|---|---|
| 100,000 to 1 million | $8.00 |
| 1 million to 10 million | $7.00 |
| 10 million to 50 million | $6.50 |
| Above 50 million | $6.00 |
Three worked months
On the Core plan, with the first 100,000 units included:
| Monthly units | Band 1 | Band 2 | Band 3 | Total |
|---|---|---|---|---|
| 1,000,000 | 900,000 at $8 = $72 | $29 + $72 = $101 | ||
| 10,000,000 | $72 | 9,000,000 at $7 = $630 | $29 + $702 = $731 | |
| 50,000,000 | $72 | $630 | 40,000,000 at $6.50 = $2,600 | $29 + $3,302 = $3,331 |
On Pro, add $170 to each total for the three-year data access. In run terms, using eleven units per run: about 90,000 runs for $101, 900,000 runs for $731, and 4.5 million runs for $3,331.
That is cheap per run compared with per-trace pricing at $5 per thousand, and the graduation keeps it from running away. The cost is in the shape: the bill moves with how richly you instrument, so a team that adds a scoring pass or a more detailed trace sees the number move, and a team on Core sees its data roll off after 90 days unless it pays for Pro.
Where the bill surprises people
- Scores are units. Running an LLM-as-judge evaluation on every trace doubles the trace count in units before the observations are counted.
- Rich traces cost more than thin ones. Instrumenting every tool call and retrieval step is the point of tracing, and each one is a unit.
- Data access, not retention, is the tier difference. Core keeps 90 days. If you need to look at a trace from last quarter, that is Pro.
- Self-hosting is free but not free. The teams that run Langfuse at volume usually self-host, which moves the cost from the invoice to the Postgres and ClickHouse instances and the engineer who looks after them.
What this buys you
Langfuse is a strong, open platform for tracing, prompt management, datasets and evaluation, with a large integration list and an honest free tier. If you want to own the data and have the people to run it, self-hosting is hard to beat on price.
What it does not do is tell you whether a scheduled agent stopped running, or page you when one does, or score the output of every run against a rubric without you building the evaluation pipeline, or give you a flat bill with no unit counting. Those are the production questions Agent Ping is built around: heartbeat and schedule monitoring, cost per customer and feature on every run, quality scoring out of the box, on a flat monthly tier with named limits. The Agent Ping vs Langfuse comparison and the migration guide cover the detail, and the LLM token cost calculator estimates the model spend inside each run, which is normally the bigger number.