In the Langfuse vs LangSmith decision, choose Langfuse if you want open-source software you can self-host for free, framework-neutral tracing built on OpenTelemetry, and simple usage-based pricing. Choose LangSmith if your application is built on LangChain or LangGraph and you want a managed product with the tightest integration, and you are happy to pay per seat. Both cover tracing, prompts, datasets and evaluations well, so the real differences are license, hosting, pricing model and ecosystem.
Both products have changed in the last year: ClickHouse acquired Langfuse in January 2026, and LangSmith has grown from a tracing tool into a wider agent platform with deployment, sandboxes and a model gateway. This comparison reflects both products as of October 2026.
Langfuse vs LangSmith at a glance
| Langfuse | LangSmith | |
|---|---|---|
| License | MIT core; enterprise directories under a separate license | Proprietary |
| Self-hosting | Free, same stack as the cloud; some add-ons need a license key | Enterprise plan add-on only |
| Free cloud tier | Hobby: 50,000 units a month, 30 days of data, 2 users | Developer: 1 seat, 5,000 base traces a month |
| Entry paid plan | Core: 29 dollars a month, unlimited users | Plus: 39 dollars per seat per month |
| Pricing model | Units: traces, observations and scores | Seats plus traces, plus usage units for other products |
| Data retention | 30 days, 90 days or 3 years depending on plan | Base traces 14 days; extended traces 180 days |
| Instrumentation | OpenTelemetry-based SDKs and many integrations | traceable decorator, LangChain auto-tracing, OpenTelemetry ingest |
| Owner | ClickHouse | LangChain |
In short, Langfuse wins on openness and cost predictability for teams, while LangSmith wins on depth of LangChain integration and on bundling tracing with agent deployment. If you have several people looking at traces, LangSmith's per-seat model adds up faster; if you emit huge volumes of spans and scores, check Langfuse's unit math carefully.
Licensing and self-hosting
This is the biggest difference. Is Langfuse open source? Yes. The core repository is MIT licensed, and the self-hosting documentation says you run the same infrastructure that powers Langfuse Cloud. A small set of enterprise add-on features live in separate directories and need a license key (self-hosting docs). After the acquisition, ClickHouse and the Langfuse team said there were no planned licensing changes and that self-hosting remains first-class (announcement).
LangSmith is closed source. Can LangSmith be self-hosted? Yes, but only as an add-on to the Enterprise plan, which you arrange through LangChain's sales team (self-hosted docs). On the cloud plans, your data lives in LangChain's US or EU cloud.
Interestingly, the two self-hosted stacks look alike. Both store traces in ClickHouse, operational data in PostgreSQL, queues in Redis, and payloads in blob storage. So the operational burden is similar; the difference is that with Langfuse you can start today without a contract.
Pricing compared
Langfuse pricing on its cloud is based on units, where a unit is any trace, observation or score you ingest. The free Hobby plan includes 50,000 units a month, 30 days of data access and two users. Core costs 29 dollars a month with 100,000 units, 90 days of data and unlimited users. Pro costs 199 dollars a month with 3 years of data access, higher rate limits and compliance reports. Extra usage is 8 dollars per 100,000 units, getting cheaper at volume, and there is a 300-dollar-a-month Teams add-on for SSO and fine-grained access control (Langfuse pricing).
Is LangSmith free? For one person, yes. The Developer plan costs nothing for a single seat and includes 5,000 base traces a month. Plus costs 39 dollars per seat per month and includes 10,000 base traces a month, unlimited seats and access to deployment features. Traces beyond the allowance are pay-as-you-go, and newer products such as sandboxes and the model gateway are billed in LangChain Standard Units at 1 dollar each (LangSmith pricing). Check the pricing page for current per-trace rates before you budget.
Retention also differs. LangSmith base traces are kept for 14 days and extended traces for 180 days; you can upgrade a trace to extended retention, which is useful for traces carrying user feedback or evaluation scores. Langfuse ties retention to the plan instead.
A practical way to compare is to export a week of real traffic and count three numbers: traces, spans per trace, and people who need access. Many spans per trace pushes Langfuse cost up; many seats pushes LangSmith cost up.
Features: tracing, prompts and evaluation
Both tools cover the same core workflow:
- Tracing: nested traces for model calls, retrieval and tool calls, with token cost and latency. Langfuse is built on OpenTelemetry and offers Python and JavaScript SDKs plus integrations for most frameworks (SDK overview). LangSmith traces LangChain and LangGraph automatically when you set environment variables, and other code through the traceable decorator, a context manager or the RunTree API (custom instrumentation). It can also ingest OpenTelemetry traces (OpenTelemetry guide).
- Prompt management: both version prompts outside your code and link traces to the prompt version that produced them (Langfuse prompt management).
- Datasets and experiments: both let you build test sets from production traces and run experiments against them.
- Evaluators: both support LLM-as-a-judge evaluators, code checks, human annotation and user feedback. Our LLM-as-a-judge guide explains how to calibrate these so the scores mean something.
Where they diverge is scope. LangSmith now positions itself as an agent engineering platform, bundling deployment of LangGraph agents, sandboxes and an LLM gateway alongside observability. Langfuse stays focused on observability, prompts and evaluation, and expects you to deploy your app elsewhere.
Which one fits your stack?
- Built on LangChain or LangGraph? LangSmith is the path of least resistance and gives you the richest view of graph execution. Langfuse also has a LangChain integration, so this is a preference, not a hard requirement.
- Using raw provider SDKs, LlamaIndex, Vercel AI SDK or your own code? Langfuse's OpenTelemetry base is a better fit, and you can also send data through tools like OpenLLMetry.
- Traces must stay on your infrastructure? Langfuse, unless you are already negotiating an enterprise contract with LangChain.
- Large team, modest traffic? Langfuse's flat plans with unlimited users are usually cheaper.
- Solo developer? Both free tiers are enough to get started; LangSmith's 5,000 traces disappear quickly in load tests, while Langfuse's 50,000 units go further but count every span.
Pros and cons
Langfuse pros: MIT-licensed core, free self-hosting, unlimited users on paid cloud plans, framework-neutral and OpenTelemetry-native, cloud regions in the US, EU and Japan.
Langfuse cons: running it yourself means operating ClickHouse, PostgreSQL, Redis and blob storage; enterprise access controls cost extra; long-term roadmap now depends on ClickHouse's priorities.
LangSmith pros: the deepest LangChain and LangGraph integration, one vendor for tracing, evaluation and deployment, polished hosted experience.
LangSmith cons: closed source, self-hosting only on Enterprise, per-seat pricing, short 14-day retention for base traces.
What about Arize Phoenix and other alternatives?
If neither fits, Langfuse vs Arize Phoenix is the next comparison people make. Arize Phoenix is free to self-host with no usage limits, runs as a single container and is built on OpenTelemetry, which makes it the lightest option to stand up. Its Elastic License 2.0 is source-available rather than open source, and Arize became part of Dynatrace on October 1, 2026. We compare the two head-to-head in Langfuse vs Arize Phoenix. Opik from Comet is another Apache 2.0 option. See our full roundup of the best LLM observability tools, and our playbook on how to evaluate LLM apps for what to do with the traces once you have them.
Migrating between them
Because both accept OpenTelemetry data, the cleanest migration path is to instrument once with OpenTelemetry and point the exporter at whichever backend you choose. You can even send traces to both during a trial. Prompts and datasets need exporting and re-importing through each tool's API, so start the trial with a small, representative dataset rather than your whole history.
FAQ
Is Langfuse better than LangSmith?
Neither is better for everyone. Langfuse is better if you want open source, free self-hosting and unlimited users. LangSmith is better if you are deep in LangChain or LangGraph and want a managed product that also deploys your agents.
Is Langfuse open source?
Yes. Langfuse's core is MIT licensed and can be self-hosted for free. A few enterprise add-on features sit in separate directories and require a license key.
Is LangSmith free?
LangSmith's Developer plan is free for one seat and includes 5,000 base traces a month. Teams need the Plus plan at 39 dollars per seat per month, plus pay-as-you-go usage beyond the included traces.
Can I use LangSmith without LangChain?
Yes. You can trace any Python or JavaScript code with the traceable decorator or the RunTree API, wrap provider clients, or send OpenTelemetry traces. LangChain is simply the case with the least setup.
Can I self-host LangSmith?
Only on the Enterprise plan, where self-hosting is an add-on arranged with LangChain's sales team. A self-hosted instance runs on ClickHouse, PostgreSQL, Redis and blob storage.
Who owns Langfuse now?
ClickHouse acquired Langfuse in January 2026. The team said there are no planned licensing changes and that self-hosting remains a first-class option.