Positioning
Langfuse sits in the observability and eval slot of the LLM stack with a strong open-source story. The MIT-licensed core covers traces, prompt management, datasets, and scoring; the enterprise modules live in an ee/ folder and require a license. Self-hosting via Docker or Kubernetes is documented and supported, not an afterthought, and the project shipped at a very high release cadence through 2025 and 2026 1. ClickHouse acquired Langfuse in January 2026; the announcement says Langfuse stays open source and self-hostable with no planned licensing changes and that Langfuse Cloud keeps running as is 2. The MIT license file now names ClickHouse, Inc. as copyright holder 3.
The product lists 100+ library, framework, and provider integrations, is built on OpenTelemetry 1, runs LLM-judge evaluators on traces, and exposes a Scores API that lets external evaluators (Ragas, DeepEval, your own) push results back to the same trace.
Langfuse v4, announced in August 2026, moves to an observation-centric data model: one append-only observations table instead of separate trace and observation tables joined at query time. It needs Python SDK 4.7.0 or newer and JS/TS SDK 5.4.0 or newer, because those versions copy trace-level attributes such as user and session ids onto every observation. Self-hosters can upgrade on their own schedule (v3 gets security patches through January 2027). Langfuse Cloud runs v3 and v4 side by side until November 16, 2026 and is v4-only after that, so existing Cloud projects should run the Migration Assistant and update SDKs, evaluators, and API calls before the cutover 4.
Strengths
- Self-host actually works. Docker Compose for a single VM, Helm for Kubernetes, plus guides for AWS, Azure, GCP, Render, and Railway. The stack is Postgres, ClickHouse, Redis or Valkey, and S3-compatible blob storage, with a web container and a worker container on top. Teams with data-residency requirements can run Langfuse on their own infrastructure without paying for a hosted plan.
- High velocity. v4.49.0 shipped on 2026-10-01, one of at least five releases between 2026-09-25 and 2026-10-01 5. Bug fixes and integrations land quickly.
- Open Scores model. The Scores API lets you attach numeric, categorical, or boolean scores to any trace from any source. Ragas, DeepEval, and custom evaluators all push to the same surface 6.
Limitations
- Governance features are license-gated when self-hosted. Project-level RBAC, protected prompt labels, data retention policies, audit logs, server-side data masking, UI customization, SCIM, and the organization and instance management APIs need an enterprise license key 7. LLM-as-a-judge evaluators, annotation queues, prompt experiments, and the playground have been MIT since June 2025, and basic SSO ships in the OSS version 8.
- UI is observability-first. Compared to Braintrust or LangSmith, the offline-eval workbench has been the thinner part, and most teams use Langfuse as the trace hub and run evals from notebooks or CI. The August 2026 evaluator rebuild closes some of that gap: you can test an LLM-judge or code evaluator against real observations before deploying it, map variables by clicking instead of writing JSONPath, see matching volume and estimated cost first, and grade images, audio, video, and PDFs captured in traces. Evaluators and evaluation rules also have a stable ID-based API under
/api/public/v2 9.
- Vendor-published benchmarks are vendor-published. As with every other platform, treat performance claims from the company as positioning.
Best fit
Production observability where data residency matters, hybrid architectures where Langfuse is the trace hub and CI pushes results back via the Scores API, and teams that want to keep platform costs predictable by self-hosting.
Getting started
from langfuse import Langfuse
from langfuse.openai import openai # auto-traced
langfuse = Langfuse()
# Run your app inside a span. The wrapped OpenAI call is traced under it.
with langfuse.start_as_current_observation(as_type="span", name="summarize-email"):
response = openai.chat.completions.create(
model="gpt-4o-mini",
messages=[{"role": "user", "content": "Summarize this email."}],
)
# Push an external score (from Ragas, DeepEval, or a human reviewer) onto the trace.
langfuse.score_current_trace(
name="faithfulness",
value=0.92,
data_type="NUMERIC",
)
To score a trace after the fact, langfuse.create_score(name=..., value=..., trace_id=..., data_type=...) takes a trace id you stored earlier. The Python SDK no longer has a top-level score() method. The Langfuse cookbook also documents a full RAG evaluation loop using Ragas, including dataset construction and dashboarding 10, and an error-analysis recipe that uses Langfuse as the trace store 11.
Pricing notes
The public pricing page (fetched as of 2026-10-01) lists Hobby at 29/month (100k units, 90 days, unlimited users), Pro at 2,499/month (100k units, 3 years, SCIM and audit logs) 12. Extra units on the paid tiers cost 6 per 100k above 50M units. A unit is a trace, an observation, or a score, and units created by LLM-as-a-judge evaluators and experiments count too, so evaluation traffic draws on the same quota as production traffic 13. Enterprise SSO, SSO enforcement, and project-level RBAC on Pro come from a $300/month Teams add-on. Self-host of the MIT core is free; the enterprise modules require a paid license. Confirm the current unit allowances and overage tiers on the pricing page before committing.
Alternatives
Consider LangSmith if your codebase is built on LangChain or LangGraph. Consider Arize Phoenix for an OpenTelemetry-native OSS hub. Consider Braintrust if a hosted, framework-agnostic console with strong dataset tooling is the priority and self-host is not required. For pure CI regression, pair Langfuse with DeepEval or Promptfoo.
Citations and last verified
Verified 2026-10-01 against the Langfuse pricing page, LICENSE file, GitHub release list, v4 changelog, self-hosting docs, and the Langfuse cookbook.