Braintrust if your team works evals-first: experiments, scorers and prompt iteration before anything ships. Langfuse if you want tracing first, open source and self-hosting, at a fraction of the price. That’s the answer to the question. The bigger decision sits after it, because neither one tells you whether your agent is any good.
Braintrust vs Langfuse at a glance
| Braintrust | Langfuse | |
|---|---|---|
| Built around | Evals, experiments and scorers | Tracing, with evals and prompts on top |
| Open source | Not stated on its site | Yes, MIT core |
| Self-hosting | Data plane in your cloud, UI hosted by Braintrust (Enterprise) | Free, the whole stack |
| Ownership | Independent | ClickHouse, since January 2026 |
| Free tier | 1 GB data, 10k scores, unlimited users, 14 days | 50k units, 2 users, 30 days |
| Entry paid plan | Pro: $249 per month, 5 GB, 50k scores, 30 days | Core: $29 per month, 100k units, unlimited users, 90 days |
| Billing unit | Gigabytes processed, plus scores | Units: each trace, observation or score |
| Overage | $3 per GB and $1.50 per 1k scores (Pro) | From $8 per 100k units, cheaper at volume |
| Built-in AI help | Loop: generates test cases, runs evals, iterates prompts | None comparable |
| Judge vs human agreement | No named feature | Score Analytics (beta): kappa, F1, confusion matrix |
Sources: Braintrust pricing, Braintrust self-hosting, Langfuse pricing, Langfuse self-hosting, Langfuse Score Analytics, ClickHouse acquires Langfuse. Checked 9 Nov 2026.
Where they’re the same
Both capture traces from your agent. Both run LLM-as-judge and code scorers, store datasets, run experiments against them, and give your team a way to review outputs by hand. Both have a playground for prompts. If you lined up the feature lists you’d find more overlap than difference.
So pick on the three things below.
Where they differ
Where the product starts. Braintrust starts from the eval: you write a scorer, run an experiment, compare it to the last one, and ship when the numbers say so. Tracing feeds that loop. Langfuse starts from the trace: you instrument the agent, watch what it does in production, and add evals and prompt management around it. Neither is wrong. They suit teams that start in different places.
Pricing shape. Braintrust charges for gigabytes of processed data and for scores. That means a support agent with long tool outputs costs more than a chatbot with short replies, even at the same traffic. Langfuse counts units: a trace, an observation or a score each counts once, whatever its size. On Langfuse, the bill follows traffic. On Braintrust, it follows traffic, trace size and how much you score.
Who holds your data. Langfuse is MIT-licensed. You can run all of it yourself for free, as long as you’re happy to operate Postgres, ClickHouse, Redis and blob storage. Braintrust’s self-hosting puts the data plane in your cloud while Braintrust runs the interface, and it sits on the Enterprise plan. On retention, Braintrust’s self-serve plans keep data for 14 or 30 days. Langfuse keeps it for 30 days, 90 days or 3 years, depending on plan.
Where both stop
You don’t buy an evals tool to have scores. You buy it to make the agent better. On that job, both stop in the same three places.
Agreement is a check, not a method
Langfuse now reports how often a judge agrees with your human scores, in beta. Braintrust’s own guides describe comparing the two by hand. Either way, that’s a check on labels you’ve already made. Neither picks which traces your expert should label, so a small budget covers the failures that matter. Neither turns a few dozen labels into a corrected rate for the whole population, with error bars. And neither keeps that up as your judges and data drift. Without it you have a pile of labels, not a measurement. More on that in what makes a good LLM judge.
The outcome isn’t in either tool
The outcome your agent exists for usually lives somewhere else. One customer runs a sales agent whose success is measured in sends. Sends live in their product database, not in any trace. Evals without outcomes are a fine starting point. They give you a well-behaved agent. If you want to move the number the business cares about, you need to know which evals predict it, and neither tool joins scores to outside outcome data.
A score doesn’t say what to change
A judge tells you 18% of conversations fail. Change the prompt, swap the model, add a tool, split the agent? Braintrust’s Loop helps you iterate on a prompt against a scorer, which is useful. It doesn’t tell you whether that scorer was the right thing to move. The path from a score to a tested change that matters is still yours to build.
I wrote the longer version of this argument in LangSmith vs Langfuse.
Which should you choose?
- Your team iterates on evals before it ships: Braintrust. The experiment workflow is the centre of the product. Budget for $249 a month once you outgrow the free plan, and check your trace sizes first.
- You want tracing first, flat pricing and open source: Langfuse. It’s the cheaper option at almost any volume, and you can leave whenever you like.
- You need to self-host everything: Langfuse. Braintrust keeps the interface on its side.
- You own whether the agent is any good: keep whichever tool you have and add an agent quality layer on top. That’s what TwoTail is. It connects to Langfuse with a pasted key and no code change, and takes OpenTelemetry from anything else.
Disclosure: I build TwoTail, so weigh the last bullet accordingly. Other comparisons: Arize Phoenix vs Langfuse, LangSmith vs Langfuse pricing.