Langfuse vs. Braintrust

This guide outlines the key differences between Langfuse and Braintrust to help engineering teams choose the right AI engineering platform. All Braintrust facts were checked against public Braintrust sources in August 2026.

TL;DR

Choose Langfuse if you want:

1. Open source and full control. MIT-licensed self-hosting of the entire application on every tier, with data in a database you can query directly.

2. Production observability at scale. Complete traces at high volume on ClickHouse, with dashboards, alerts, and full-text search. Observability is the foundation; evals, prompts, and datasets layer on the same data.

3. Predictable, count-based pricing. A trace, observation, or score is one unit: cost follows request count, not payload size, and scores are not a separate meter. Pro includes 3-year retention.

Choose Braintrust if your team is evaluation-first and wants a vendor-managed workbench for reviewing experiments, with AI-assisted analysis built in.

---

Open source and distribution

The core difference is the distribution model. Langfuse is open source, and self-hosting is a first-class deployment mode: core features ship in the MIT distribution with no usage limits. Braintrust is proprietary: its SDKs and the autoevals evaluator library are open, but the platform, control plane, and storage engine are not open source.

Langfuse self-hosting runs the entire application in your environment; Langfuse Cloud is the managed alternative. Braintrust always operates the control plane (product UI, authentication, organization metadata) itself. On the Enterprise plan, the data plane (traces, datasets, prompts) can run in your AWS, GCP, or Azure account, operated either by Braintrust with access to your cloud (BYOC) or by your team, as closed software it cannot inspect or patch (self-hosted).

That model fits teams whose requirement is keeping payloads in their own VPC. If the UI, authentication, and control metadata must also stay in your environment, or you need air-gapped operation, only Langfuse self-hosting satisfies it.

---

Storage: ClickHouse vs Brainstore

Both platforms re-architected storage for agent-scale trace data.

Braintrust built Brainstore, a Rust engine that writes spans to object storage and indexes them with the open-source Tantivy library. It runs inside the data plane, so in hybrid deployments it lives in your cloud account and follows Braintrust's release cadence.

Langfuse runs on ClickHouse, an open-source analytical database proven at petabyte scale across thousands of companies since 2016, and works directly with its core team as part of ClickHouse. Every LLM call, tool execution, and agent step is written once to a wide, immutable observations table (v4 GA; details).

Brainstore is queried through Braintrust's APIs and BTQL. ClickHouse speaks standard SQL with broad tooling and operational expertise you can hire for: run and query it directly when self-hosting; on Cloud, high-volume reads go through the Observations and Metrics APIs.

---

Pricing

The two meters are fundamentally different. Langfuse counts events: 1 unit = 1 trace, observation, or score. Braintrust runs four meters: processed data per GB (inputs, outputs, prompts, metadata, spans, datasets, and attachments all count), a per-score charge, a retention charge, and monthly model credits for AI features like Topics and built-in models, billed per token past the included credit (Braintrust pricing).

Worked example

500,000 traces/month · 10 spans per trace · 5 KB per span · 1 score per trace · 12-month retention · 5 users.

Public list prices. Edit the inputs yourself: Langfuse vs Braintrust pricing model.

1. Scores drive the Braintrust usage line: $675 of the $735 is score charges at 1 score per trace, and $60 is processed data. Every score is metered, so scaling online evaluation scales the bill directly. On Langfuse, a score is one unit, same as a trace.

2. *The retention line is a steady-state estimate: 25 GB/month held for 12 months, billed at the published $0.50/GB/month on every GB older than the included 30 days (formula in the sheet). Braintrust Pro caps retention at 180 days, so a full year in practice requires an Enterprise contract. Langfuse retention has no separate meter.

3. Which model is cheaper depends on trace shape. Payload-light, score-light workloads with short retention can be economical on Braintrust; large payloads, high score density, long retention, or simply high request volume favor Langfuse. The sheet takes span size and score density as inputs so you can find the crossover for your workload.

See pricing · Start Free · Talk to us

---

Observability and monitoring

Langfuse starts from production tracing: complete traces at high volume with token and cost tracking and Pulse outlier detection on the same observations table, and SDK-level sampling to cap ingestion volume at the source.

Braintrust's Observe covers trace inspection, monitoring, and Topics, a daily pipeline that classifies logs by intent, sentiment, and issues. Topics' model calls are served from Braintrust-hosted endpoints in every deployment mode, including BYOC and self-hosted, with zero data retention at the model host. Trace text crosses your deployment boundary even when the data plane runs in your cloud. Telemetry arrives via SDK wrappers or OTLP traces. Langfuse ingests OpenTelemetry natively and ships 100+ framework and provider integrations, with the SDKs sending telemetry asynchronously outside the request path.

---

Evaluation and experiments

Both platforms cover offline and online evaluation end to end; the differences are workflow packaging and openness.

Braintrust has persistent experiment baselines, row-level regression sorting, repeated trials with statistical summaries, Loop (an in-product AI agent that analyzes traces and builds datasets and scorers), Topics (daily semantic classification of logs), and generated custom review views. Teams that live in experiment review get a mature workbench.

Langfuse evaluation is built for flexibility: LLM-as-a-judge and deterministic code evaluators run online and in experiments, on the same observations table production writes to, with rules controlling targeting, sampling, and cost. The loop from production to improvement closes inside one project: add an observation to a dataset in one click, run an experiment against it, and route outputs to annotation queues for human review. It is customizable down to the scorer: judges run on any connected model, code evaluators run your own Python or TypeScript, and Braintrust's open-source autoevals scorers run unchanged via a built-in converter. It is also an area of visible investment: multi-modal datasets, tool-call evaluation, and reusable evaluators with cost preview shipped in the past three months, and better evals and experiments is a named focus area on the public roadmap.

---

Open platform and agentic access

Langfuse is API-first, and both platforms ship agent-facing surfaces (CLI, MCP server, agent skills). The difference is what sits underneath: because the Langfuse CLI and MCP server are part of the OSS product, agentic access works on any self-hosted instance, including the free tier, against a database you control.

---

Enterprise security

Langfuse publishes SOC 2 Type II, ISO 27001, GDPR, and HIPAA and signs a Business Associate Agreement (BAA) from the Pro plan in the dedicated HIPAA Cloud region. Braintrust publishes SOC 2 Type II, GDPR (DPA), and HIPAA, with BAAs available on Enterprise plans; ISO 27001 is not listed on their security page. Langfuse is used by 21 of the Fortune 50 and 129 of the Fortune 500.

---

Which should you choose

Choose Langfuse if your deciding constraints are infrastructure and scale: the whole platform needs to run in your environment, you want to query trace data directly in SQL, or production volume makes a per-request meter the safer budget. The full evaluation stack (judges, code evaluators, experiments, annotation) comes with it in the same project.

Choose Braintrust if your deciding constraint is a vendor-managed review workbench: packaged experiment baselines, repeated trials with statistical summaries, and AI-assisted analysis through Loop and Topics, with sensitive payloads optionally kept in your VPC on Enterprise.

---

Why teams move away from Braintrust

The same reasons recur in migration conversations. Cost leads, and it is the meter design more than the list price: the worked example lands at $1,122/month on Braintrust Pro vs $621 on Langfuse Pro.

• Size-metered ingestion. Cost follows payload size on the GB meter, not request count, so high-volume applications with many simple traces feel it hardest.
• Scores are their own meter. Every score is a metered charge, so evaluation coverage becomes a budget decision (worked example).
• Retention is short and metered. Keeping data past the included window bills per GB monthly and caps at 180 days on Pro, so multi-year lookback needs an Enterprise contract.
• Eval-first packaging. Teams whose primary job is production observability end up paying for an evaluation-centric package they use as a logging system; Langfuse is built the other way around.
• Getting data out takes work. Bulk export from Braintrust means cursor pagination in 1,000-record pages or the CLI; automated cloud-storage export is Enterprise-gated, and there is no direct database access. Langfuse pairs high-volume read APIs and blob-storage export with direct SQL on ClickHouse when self-hosting.
• Full self-hosting is not offered. Braintrust's self-hosted option covers the data plane only, requires Enterprise, and keeps a runtime dependency on Braintrust's cloud for UI and auth. When something breaks in closed software, only the vendor can diagnose and fix it. On Langfuse, running the whole application yourself is a first-class deployment mode, open source down to the storage engine.

---

Why teams choose Langfuse

A pattern repeats across production adopters: start where governance requires, often self-hosted, then grow into the platform.

• SumUp saved 30% of external BPO cost by deflecting 50% of support conversations to AI (Ana Casado). They started self-hosted for the PoC, then moved to Langfuse Cloud for production.
• Canva runs framework-agnostic raw LLM clients: Java via OpenTelemetry, Python via the Langfuse SDK. Non-technical users contribute domain expertise through judges. They self-hosted first, then moved to Cloud.
• Merck runs about 80 GenAI use cases on Langfuse with 200+ people building on it. They self-host for data sovereignty, provision via API, and cited the public roadmap and shipping velocity in the selection.
• Khan Academy built a custom Go client on the open API, serving 100+ users across 7 product and 4 infrastructure teams.

---

Switching from Braintrust

Teams that switch rarely migrate everything. They run both platforms in parallel, repoint live instrumentation, port the durable assets, validate at production scale, then turn Braintrust ingestion off. Most see their first traces in Langfuse the same day.

• Instrumentation: swap wrap_openai/@traced for the Langfuse equivalents, or, if you emit OpenTelemetry, point the exporter at the OTLP endpoint. Both platforms can trace the same app side by side during the transition.
• Datasets, prompts, and evaluators: export datasets via the Braintrust API and import through the Langfuse SDK; your autoevals scorers run unchanged. The step-by-step migration guide covers each asset.
• Trace history: stays readable in Braintrust for its retention window while live data builds up in Langfuse from day one.

Talk to us if you want help planning a migration.

---

Start free: Cloud or self-host

Start on Langfuse Cloud (Hobby: 50k units/mo, no credit card) or self-host the MIT-licensed stack. You can explore a populated example project before instrumenting anything.

---

FAQ

Is Langfuse the best Braintrust alternative?

Yes, if you want an open-source platform where production observability, evaluation, and prompt management run on infrastructure you can inspect, query, and self-host. See Which should you choose.

Is Braintrust open source?

No. The Braintrust platform, control plane, and Brainstore storage engine are closed source; the SDKs and the autoevals evaluator library are open source. Langfuse is MIT-licensed and self-hostable.

Can I self-host Braintrust?

Partially. On the Enterprise plan, the data plane (traces, datasets, prompts) can run in your AWS, GCP, or Azure account, while the control plane (UI, authentication, platform management) is always operated by Braintrust. Langfuse self-hosts the entire application on every tier; see Open source and distribution.

Is Langfuse only an observability tool?

No. Tracing, prompt management, LLM-as-a-judge and code evaluators, datasets, experiments, annotation queues, dashboards, and alerts ship as one integrated platform in the same project. See the clarifications page for dated, linked facts.

Can product managers work in Langfuse without engineering handoffs?

Yes. Non-engineers iterate in the playground, edit and deploy prompt versions via labels that applications fetch at runtime (no redeploy), run experiments from the UI, and review outputs in annotation queues. Read the Canva case study: a marketing manager set up a dedicated LLM-as-a-judge evaluator without engineering help.

Can I keep my autoevals scorers?

Yes. autoevals is an open-source library, and Langfuse's experiment SDKs include converters that run autoevals scorers unchanged. See the migration guide.

How does Braintrust pricing compare to Langfuse?

The meters differ: Langfuse counts traces, observations, and scores as units (Core starts at $29/month with unlimited users); Braintrust bills processed data per GB, plus per-score charges, retention, and model credits for AI features. On the rates in our editable public model, 500k traces / 10 spans / 5 KB per span / 1 score per trace / 12-month retention is $621 Langfuse Pro vs $1,122 Braintrust Pro at published rates (Braintrust Pro caps retention at 180 days, so a full year in practice requires Enterprise). Which is cheaper depends on payload size, score density, retention, and volume; details in the worked example.

How do I migrate from Braintrust to Langfuse?

Run both in parallel, swap the SDK wrappers or repoint your OpenTelemetry exporter, import datasets via API, and keep your autoevals scorers. Most teams see first traces the same day; history stays in Braintrust. Follow the migration guide or see Switching from Braintrust.

---

---