Both are LLM and agent evaluation platforms with tracing, offline experiments, online scoring and prompt versioning. Langfuse is the open-source option you can run yourself for free; Braintrust is a hosted platform organised around experiments and an in-product analysis agent. For most teams the licence and hosting question decides it before any feature does.[1][4]
Langfuse is open source (MIT, apart from licensed add-ons) and free to self-host; ClickHouse acquired it in January 2026. Braintrust is a hosted, proprietary platform built around experiments, with self-hosting of its data plane on Enterprise only. Both trace agents, score production traffic and version prompts, and neither can block a live agent action.
| Decision factor | Braintrust | Langfuse |
|---|---|---|
| Ownership | Independent; $80M Series B in February 2026 | Acquired by ClickHouse in January 2026; roadmap and licence unchanged, per Langfuse |
| Licence | Proprietary platform; SDKs and Autoevals are open source | MIT, except add-on features in the ee folders |
| Self-hosting | Enterprise only, and only the data plane; Braintrust runs the control plane | Free, via Docker or Kubernetes |
| Centre of gravity | Experiments: dataset, task, scorers, compare | Traces: instrument, observe, then evaluate |
| Online evals | Yes, sampled scoring rules on production logs | Yes, evaluator rules with sampling, plus backfills over past data |
| Gateway | Hosted gateway in public preview, no SLA | None; traces are exported out of the request path |
| Free tier | Starter: 1 GB processed data, 10K scores, 14-day retention | Hobby: 50k units a month, 30 days of data; self-hosting free |
| First paid tier | Pro: $249 a month | Core: $29 a month |
| Capability | Braintrust | Langfuse | Prefactor |
|---|---|---|---|
| Observe | |||
| LLM call tracing | ✓ | ✓ | ✓ |
| Multi-step agent tracing (tool calls, nested spans) | ✓ | ✓ | ✓ |
| Cost tracking | ✓ | ✓ | ✓ |
| OpenTelemetry support | ✓ | ✓ | ✓ |
| Evaluate | |||
| Offline evals (datasets, experiments) | ✓ | ✓ | ✓ |
| Online evals on production traffic | ✓ | ✓ | ✓ |
| LLM-as-a-judge scoring | ✓ | ✓ | ✓ |
| Code-based evaluators | ✓ | ✓ | ✓ |
| Human review and annotation | ✓ | ✓ | ✓ |
| Prompt versioning and playground | ✓ | ✓ | — |
| Act on findings | |||
| Runtime controls on live agent actions | — | — | ✓ |
| Gateway or proxy in the request path | ✓ | — | — |
| Licence and deployment | |||
| Open source | Partial | ✓ | Partial |
| Self-hosting | Partial | ✓ | Partial |
Feature lists for the two products now overlap almost completely: both trace multi-step agents, run experiments on datasets, score production traffic with LLM judges or code, queue runs for human review and version prompts. What still differs is where the software runs and who owns the code.
Langfuse publishes its platform under the MIT licence and documents a free self-hosted path on Docker or Kubernetes, backed by Postgres, ClickHouse, Redis and object storage. A small set of add-on features sits in licence-keyed ee folders. If your traces cannot leave your network, Langfuse is the option that works without a sales conversation.
Braintrust is hosted. Its Enterprise plan lets you run the data plane in your own cloud, but Braintrust always operates the control plane: the UI, logins and metadata. If a compliance review requires everything inside your own network, check that detail before you trial it.
Braintrust's product starts from the experiment: you bring a dataset, define a task and scorers, run variants and compare. Its Loop agent reads results and suggests changes, and since September 2026 Patterns and Topics group production behaviour for review (several of these are in public preview).
Langfuse starts from the trace. You instrument, look at what production is doing, then build evaluations from it. Since April 2026 its experiments run as a top-level feature against datasets, production traces or local data, and since September evaluators can backfill scores over up to six months of past observations.
Teams that are pre-launch and iterating on prompts tend to find Braintrust's loop faster. Teams already in production that need to see what is happening first tend to start with Langfuse. Both now cover the other motion.
The two price on different units, so compare on your own volumes. As of 20 October 2026, Langfuse charges per unit ingested (traces, observations and scores): Hobby is free for 50k units a month, Core is $29 a month for 100k units, then $8 per 100k. Braintrust charges a platform fee plus processed data and scores: Starter is free with 1 GB and 10K scores, Pro is $249 a month with 5 GB and 50K scores, and both include unlimited users.
Self-hosted Langfuse removes the subscription but adds the infrastructure you run. Check Braintrust's pricing and Langfuse's pricing before you decide; both changed plans this year.
Both platforms score runs after they happen. Langfuse's own documentation says run-time protection comes from separate libraries and that Langfuse evaluates those measures after the fact. Braintrust's online scoring, alerts and Patterns also judge traces after logging. Neither can hold a refund for approval, block a tool call that breaks a rule, or stop a run that has gone wrong.
For many teams that is fine: evaluation after the fact is how they find and fix problems. If an agent takes actions with real consequences, such as payments, record changes or messages to customers, check how you will act on a failing score before the next run makes the same mistake.
We sell the layer this section describes. Read it with that in mind.
Stated plainly, since Prefactor publishes this page: Prefactor is an agent evaluation platform that scores each step of a run, not only the final answer, and can act on what it finds through runtime policies, approval holds and a kill switch scoped to a run, an agent or a team. That is the gap both Braintrust and Langfuse leave open. Prefactor has no prompt playground or prompt registry and its platform is hosted, so if prompt iteration or free self-hosting is what you need, Braintrust or Langfuse is the better pick. See the full comparison of 15 platforms.
Reviewed against public product and documentation pages on October 20, 2026. If a vendor has changed a feature, product name, or positioning since then, send a correction and we will update it. Numbered source links in the page body point to the ordered sources below.
Prefactor scores each step of every agent run and can hold or stop an action when a score says it should not happen. Start on the free tier: 5,000 spans a month, no credit card required.
Prefactor helps teams observe, evaluate, and improve their AI agents in production, across every framework and provider.