Langfuse vs LangSmith: An Independent Comparison

TL;DR

• Langfuse core is MIT-licensed and self-hostable free at unlimited traces, users and projects. Nine named features sit behind a paid Enterprise key.
• LangSmith self-hosting is Enterprise-tier only. Developer is $0 (1 seat, 5k base traces/mo); Plus is $39/seat/mo (10k base traces/mo).
• Langfuse was acquired by ClickHouse in January 2026. Self-hosted Enterprise is now bundled with a ClickHouse commercial plan.
• Both tools bill you more for running evaluations. Neither says so on its comparison page.

Pick Langfuse if self-hosting matters: the core is MIT-licensed and free to run at unlimited volume. Pick LangSmith if you already live in LangGraph and want managed agent deployment next to your tracing — but self-hosting is Enterprise-only. Both now accept OpenTelemetry, which weakens lock-in on either side.

What is Langfuse, and what is LangSmith?

Langfuse is an open-source LLM engineering platform: tracing, prompt management, evaluations, datasets and a playground. Its own README describes it as helping teams "develop, monitor, evaluate, and debug AI applications." It runs on Postgres, ClickHouse, Redis and S3-compatible storage.

LangSmith is LangChain's commercial platform. As of 2026 it is no longer just observability and evals — it now bundles Deployment, Sandboxes, an LLM Gateway, an autonomous debugging product called Engine, and a no-code agent builder called Fleet.

That scope difference is the real story. LangSmith is becoming an agent platform. Langfuse is staying an observability and evaluation layer.

One fact neither vendor's comparison page mentions: ClickHouse acquired Langfuse in January 2026. The Langfuse repository confirms it — "since January 2026 we're part of ClickHouse."

Is Langfuse open source? Yes, with a precise caveat

Langfuse's repository README states it exactly: "This repository is MIT licensed, except for the ee folders."

So the core is genuinely MIT. Not Business Source License, not "open core with a three-year delay." MIT. The repo carries 32.4k stars and 3.5k forks.

The self-hosted OSS build is explicitly uncapped. Langfuse's self-host pricing page says you can "self-host all core Langfuse features for free without any limitations," and lists included usage, users and projects as "Unlimited."

Nine features require a paid Enterprise license key, per Langfuse's license-key documentation:

• Project-level RBAC roles
• Protected prompt labels
• Data retention policies
• Audit logs
• Server-side data masking
• UI customization
• Organization creators
• Org Management API and SCIM
• Instance Management API

Worth noting what is not on that list: Enterprise SSO and organisation-level RBAC are included in the free build. That is unusually generous — most open-core vendors put SSO behind the paywall.

The docs are blunt about the free tier: "All core Langfuse features and APIs are available in Langfuse OSS (MIT licensed) without any limits."

What LangSmith offers for self-hosting

Self-hosted and hybrid deployment appear only on the Enterprise plan, at custom pricing. Developer and Plus are cloud-only, hosted in LangChain's US or EU regions.

If running the platform inside your own VPC is a hard requirement, LangSmith means a sales call. Langfuse means a Helm chart.

What does each one cost as of September 2026?

The pricing models are not comparable like-for-like, and that matters.

Langfuse charges per unit. Langfuse's pricing page defines it as arithmetic: Units = Count of Traces + Count of Observations + Count of Scores. Overage is $8 per 100k units, sliding to $6 per 100k above 50M units. Seats are free on every paid tier.

LangSmith charges per seat plus usage. Beyond included traces, LangChain's pricing page meters in LangChain Compute Units ($1.50 each) and LangChain Storage Units ($1.00 each). LangSmith no longer publishes a flat per-1,000-trace price; it publishes a calculator. Budget accordingly — a $39/seat headline on a ten-person team is $390/month before a single trace is billed.

A retention detail worth catching: LangSmith base traces retain for 14 days, extended traces for 180 days. LangSmith's usage and billing docs state that from 14 September 2026 the SaaS maximum dropped to 180 days. Its pricing-page FAQ still says 400 days. The docs are newer; treat 180 as current and confirm before you build a compliance case on it.

The billing behaviour both tools share, and neither advertises

Running evaluations increases your bill on both platforms. This is the single most useful thing to know before you commit.

Langfuse's billable-units documentation states it plainly: any trace, observation or score counts as a billable unit "whether it is sent by your application or created by Langfuse features such as LLM-as-a-Judge, Annotation Queues, or experiments."

LangSmith's version is subtler and costs more. Online evaluators and automation rules silently auto-upgrade a base trace to extended retention, at the higher rate. A thread-level rule upgrades every trace in the matched thread. You can opt out per evaluator, but it is on by default.

LangSmith at least defends the design in its docs, arguing that traces with meaningful interaction deserve a higher rate. Fair enough. But if you turn on aggressive online evals and do not read that page, your invoice will surprise you.

Two other LangSmith limits worth knowing: a Developer account with no payment method caps at 5,000 traces/month, and any single trace is capped at 25,000 runs.

Do you need LangChain to use LangSmith?

No. This is the most out-of-date criticism of LangSmith, and it deserves retiring.

LangSmith accepts standard OTLP at api.smith.langchain.com/otel and maps OpenInference, TraceLoop and Logfire attributes as well as the GenAI semantic conventions. You can trace a plain OpenAI call with the vanilla OpenTelemetry SDK and nothing from LangChain.

It goes further: LANGSMITH_OTEL_ONLY="true" sends traces only to a third-party endpoint. LangChain ships a documented escape hatch from its own product.

Langfuse is equally open, with one concrete limitation. It receives OTLP on /api/public/otel over HTTP/JSON and HTTP/protobuf — gRPC is not supported. If your collector fleet is standardised on OTLP/gRPC, that is a real integration task, not a footnote.

When the answer is neither

If all you want is LLM spans landing in the observability stack you already pay for, buy neither.

Both platforms are now OTLP endpoints with a UI on top. If you run an OpenTelemetry Collector, you can fan the same span stream out to the general-purpose APM tools you already run alongside — or instead of — either tool. LangSmith's docs include the collector config for exactly this.

Adopt Langfuse or LangSmith when you want the things generic APM does not give you: prompt versioning, datasets, LLM-as-a-judge scoring, and human annotation queues. Those are the genuine reasons to add a specialist tool. Trace visualisation alone is not.

Note also that observability is a different layer from routing. If you are choosing what sits in front of your model calls rather than what records them, that is a gateway decision, and a separate one from which engine serves the model.

What this means for you

If you're a solo dev or a small team: Langfuse Hobby (50k units, free) beats LangSmith Developer (5k traces, 1 seat) by an order of magnitude on volume. Not close.

If you run a platform team with a data-residency mandate: Langfuse, self-hosted. MIT, unlimited, in your VPC, with SSO included. This is the clearest win either tool has.

If you're building on LangGraph and want managed deployment: LangSmith. The integration is tighter and Deployment, Sandboxes and Gateway have no Langfuse equivalent. You are buying an agent platform, not an observability tool.

If you're a regulated enterprise: Both work. Langfuse Pro ($199/mo) gets you SOC 2 Type II and ISO 27001 reports plus a HIPAA-ready region at a published price. LangSmith Enterprise requires a sales conversation.

One caution on Langfuse: post-acquisition, self-hosted Enterprise is bundled with ClickHouse Cloud, BYOC or Private, and Langfuse pricing is "additive to your ClickHouse commercial plan." The MIT core is unaffected and the team has committed to keeping it open. But the commercial path now runs through ClickHouse. Price it as a ClickHouse relationship, not a standalone line item.

Frequently Asked Questions

What is Langfuse used for?

Langfuse is used to trace, debug, evaluate and monitor LLM and agent applications. Teams use it for distributed tracing of multi-step agent runs, versioned prompt management, LLM-as-a-judge scoring, dataset collection and human annotation. It integrates through native Python and JS/TS SDKs or any OpenTelemetry-compatible instrumentation.

Is Langfuse open source?

Yes. The repository is MIT-licensed except for the ee folders, per its README. Self-hosted Langfuse OSS is free with unlimited traces, users and projects. Nine features — including audit logs, SCIM, data retention policies and project-level RBAC — require a paid Enterprise license key. Enterprise SSO is free.

Is LangSmith free?

Partly. The Developer plan is $0 for one seat with 5,000 base traces per month, and caps at 5,000 traces monthly if no payment method is on file. Beyond that, Plus costs $39 per seat per month with 10,000 base traces included, then usage-based billing in compute and storage units.

What Langfuse alternatives exist?

Beyond LangSmith: Arize Phoenix, Opik, Braintrust, LangWatch and Helicone. Because Langfuse ingests standard OTLP, generic observability backends like Datadog, Honeycomb and Grafana are also viable if you only need trace visualisation and not prompt management, datasets or evaluation scoring.

Which LLM observability tool has the lowest lock-in?

Langfuse self-hosted, on licensing grounds: MIT core, your infrastructure, no vendor dependency. On the data layer both are comparable, since each ingests and re-exports OpenTelemetry. Running an OpenTelemetry Collector in front of either keeps your instrumentation portable and lets you switch backends without touching application code.

Editor's note — sources: Licensing, pricing and billing claims verified against Langfuse's license-key documentation, the Langfuse GitHub repository, Langfuse's pricing page, Langfuse's billable-units documentation, LangChain's pricing page and LangSmith's usage and billing docs, all checked 24 September 2026. Two deliberate omissions: LangSmith's per-1,000-trace dollar rates are widely quoted by aggregator sites, but the live pricing page has moved to an LCU/LSU model and no longer publishes a flat per-trace figure, so none is stated here. And LangSmith's pricing-page FAQ contradicts its own docs on maximum retention (400 days versus 180); the discrepancy is flagged in the text rather than resolved silently. Edgewisely has no commercial relationship with either vendor.