Key Takeaways
- The one-line difference — Langfuse starts from traces and adds evals. Braintrust starts from evals and adds traces. Both now cover tracing, datasets, experiments, LLM-as-judge scoring and prompt versioning, but each is noticeably better at the job it started with.
- Pick Langfuse when — You want open source, need traces in your own infrastructure without an enterprise contract, or want the cheapest path to production tracing (free self-host, Cloud from $29 a month).
- Pick Braintrust when — Evaluation is the bottleneck: you change prompts and models weekly and need fast experiment comparisons, rich scorers and a playground on production data, and you are fine with SaaS (BYOC on Enterprise).
- Ownership — Langfuse joined ClickHouse in January 2026 and committed to unchanged licensing and self-hosting. Braintrust is independent after an $80M Series B in February 2026.
At a glance
Langfuse is an open-source LLM engineering platform from Berlin. It began as tracing (what did my app and agent actually do, and what did it cost) and grew evals, datasets and prompt management around it. ClickHouse acquired it in January 2026. Braintrust began as an evaluation tool (did this prompt or model change make outputs better or worse) and grew production tracing around that loop. It raised an $80M Series B at an $800M valuation in February 2026.
Feature comparison
| Feature | Langfuse | Braintrust |
|---|---|---|
| Tracing and observability | ||
| Nested traces for agents | Core strength: sessions, nested spans, cost and latency per step, OpenTelemetry ingestion | Full traces including tool calls and retrieved context, stored in Brainstore |
| Production dashboards | Cost, latency, quality and usage dashboards across users and sessions | Monitoring views and online scoring; dashboards less central than experiments |
| Evaluation | ||
| Experiments | Datasets, experiment runs and comparisons | The centre of the product: side-by-side diffs, score deltas, regressions highlighted |
| Scorers | LLM-as-judge templates, custom scores via SDK, human annotation queues | Built-in scorers, open-source autoevals, LLM-as-judge, code scorers, AI assistant to draft them |
| Online (production) evals | Evaluators run on sampled production traces | Online scoring on production logs |
| Prompts | ||
| Prompt management | Versioned prompts with labels (production, staging), fetched by SDK, cached client-side | Version-controlled prompts tied to the playground and experiments |
| Playground | Playground for prompts and model settings | Playground that runs prompts against datasets and production data with scores inline |
| Deployment and pricing | ||
| Open source | MIT-licensed core | Commercial platform; SDKs and autoevals are open source |
| Self-hosting | Free self-hosting on ClickHouse, Postgres and app containers | BYOC (data in your cloud) on Enterprise only |
| Free tier | Hobby: 50k units a month | Starter: 1 GB processed data, 10k scores, 14-day retention |
| Paid plans | Core $29, Pro $199, Enterprise $2,499 a month; $8 per extra 100k units | Pro $249 a month (5 GB, 50k scores, 30-day retention); Enterprise custom |
Tracing: Langfuse's home ground
Langfuse's trace view, sessions and cost dashboards are the reason most teams adopt it. It ingests OpenTelemetry, has SDKs and integrations for the common frameworks, and its dashboards answer the operational questions (which users, which features, which models cost what) without extra work. Braintrust traces just as completely, including tool calls and retrieved context, but its interface treats traces mainly as raw material for evals.
Evals: Braintrust's home ground
Braintrust's experiment view is the best in the category: run a prompt or model change against a dataset, see score deltas and regressions side by side, drill into the examples that got worse, and fix them in the playground. Scorers range from the open-source autoevals library to LLM-as-judge and code. Langfuse has datasets, experiments, LLM-as-judge and human annotation queues, and they are good enough for many teams, but the loop is less tight.
Self-hosting and data control
This is often the deciding factor. Langfuse self-hosts for free from its MIT-licensed repository. Braintrust runs as SaaS on every plan; keeping data in your own cloud is its BYOC deployment, which needs Enterprise. If legal requires traces to stay in your VPC and you are not ready for an enterprise contract, Langfuse wins by default.
Decision framework
- Traces must stay in your infrastructure, no enterprise deal? Langfuse.
- Want open source? Langfuse.
- Prompts and models change weekly and quality regressions hurt? Braintrust.
- Small team, low volume, tight budget? Langfuse Cloud Core or self-host.
- Product team owns AI quality, not platform engineering? Braintrust.
Frequently asked questions
Is Langfuse or Braintrust better?
Neither is better across the board. Langfuse is better for open-source, self-hosted, observability-led teams and is cheaper to start. Braintrust is better for teams whose bottleneck is evaluation: fast experiment loops, rich scoring and a playground on production data. Many teams pick based on whether traces or evals are the first problem they need to solve.
Can I self-host Braintrust like Langfuse?
Not in the same way. Langfuse self-hosts for free from its open-source repository. Braintrust is SaaS on all plans, and keeping data in your own cloud uses its BYOC deployment, which requires an Enterprise contract.
Langfuse vs Braintrust pricing?
Langfuse Cloud is free for 50k units a month, then Core at $29, Pro at $199 and Enterprise at $2,499 a month, with $8 per 100k extra units; self-hosting has no unit charge. Braintrust Starter is free with 1 GB of processed data and 10k scores a month, Pro is $249 a month with 5 GB and 50k scores, and Enterprise is custom. At low volume Langfuse is cheaper; at scale the comparison depends on trace size versus event count, so model it on your own traffic.
Does the ClickHouse acquisition change Langfuse?
Langfuse says no: it stays open source and self-hostable, with no planned licensing changes, the same roadmap and the same Cloud contracts and support. Langfuse already stored traces in ClickHouse, so the technical fit is direct. The MIT-licensed core also means a community fork would be possible if that ever changed.
Can I use Langfuse and Braintrust together?
Yes, and some teams do: Langfuse for production tracing and cost dashboards, Braintrust for offline experiments. Both accept OpenTelemetry-style instrumentation, so the overhead is mostly in keeping datasets in sync. Most teams consolidate on one within a year to avoid paying for two trace stores.
What are the alternatives to both?
LangSmith for LangChain and LangGraph stacks, Arize Phoenix and Opik for other open-source options, Datadog LLM Observability for APM-native teams, and Promptfoo for evals in CI. See our Langfuse alternatives and Braintrust alternatives guides.
Ready to Find the Right AI Tools?
Browse our data-driven rankings to find the best AI tools for your team.