Braintrust vs Langfuse
Two trace-first evaluation platforms with different licences, different owners and very different pricing shapes. Which fits which team.
We are not in this comparison. Both of these are built around a trace of an application you wrote, which is a different job from the one we do, so this page is written as a reader rather than as a competitor.
They are closer to each other than either is to anything else in the category. Both start from the trace, both add evaluation on top, and both now have online scoring on live traffic.
The three real differences
| Braintrust | Langfuse | |
|---|---|---|
| Source | Libraries open, platform closed | MIT, with the enterprise folder excluded |
| Run it yourself | Enterprise tier only | Yes, and that is the point of it |
| Owner | Independent, Series B in February 2026 | ClickHouse, acquired January 2026 |
If being able to run the thing on your own hardware without a sales call matters, that is the whole decision and Langfuse wins it. Self-hosting Braintrust needs an enterprise contract.
Pricing is shaped differently, not just priced differently
| Braintrust | Langfuse | |
|---|---|---|
| Free tier | 1 GB data, 10,000 scores, 14 days | 50,000 units, 30 days, two users |
| Next tier | 249 dollars a month | 29 dollars a month |
| Tier above | Custom | 199 dollars, then 2,499 dollars |
| Metered on | Processed data and scores separately | Units |
| Retention at paid entry | 30 days | 90 days |
Braintrust meters scoring on its own. Scoring more of your traffic costs more there in a way it does not at Langfuse, which matters if the plan is to score everything rather than a sample.
Its free tier also allows one human review score per project. Human review is how you find out whether an automated score is any good, so that limit lands on the most useful thing to do.
Langfuse self-hosted is free, with the caveat that it wants PostgreSQL, Redis and ClickHouse running, and its own pricing page notes that enterprise self-hosting is priced on top of a ClickHouse plan.
Neither is built for a phone call
Both support audio as something attached to a trace, through framework tracing. Neither has an audio judge, acoustic metrics or telephony.
Braintrust's own cookbook scores a voice example by exact text match on the transcript. That is tracing with audio in it rather than measuring how something sounded.
How to choose
- You want to self-host without a contract: Langfuse.
- You are already a ClickHouse shop: Langfuse, and ask how the two bills interact.
- You want production observability as the primary product and evaluations derived from it: Braintrust.
- You will score a large share of traffic: compare Braintrust's score metering against Langfuse units at your real volume before deciding.
- Your conversations are phone calls and WhatsApp threads that arrive from a vendor: neither, and that is where we come in.
Everything above was read on each company's own site and documentation on 5 October 2026. Pricing and features move. Check the source before you make a decision on it, and tell us if we have something wrong.
This page as markdown