Comparison guide
Langfuse vs LangSmith: tracing, evals, pricing, and self-hosting compared
Langfuse and LangSmith both trace, evaluate, and manage prompts for LLM apps and agents. Where they differ is licensing and self-hosting, how each one bills, and how closely each is tied to the company behind it. Moda makes a product in this space, so we checked every claim against each vendor's own pages and say at the end when neither is the right fit.
Last verified against each vendor’s own docs and pricing pages · Report an error
The short answer
Pick Langfuse if you want open source, free self-hosting, or a usage-based bill with unlimited users. Pick LangSmith if you build on LangChain or LangGraph, want evaluators that score whole conversations, or want tracing, evals, and agent deployment from one vendor.
Choose Langfuse if
- You need to self-host, and want to do it for free with no usage limits.
- You want an MIT-licensed codebase you can read, audit, and fork.
- More than a few people need access: paid Langfuse plans include unlimited users, while LangSmith Plus is $39 per seat.
- You want 90 days or more of trace history without paying to upgrade traces one by one.
Choose LangSmith if
- Your agents run on LangChain, LangGraph, or Deep Agents, and you want tooling from the team that builds them.
- You want production evaluators that score a whole conversation once it ends.
- You need Go or Java SDKs.
- You want to deploy agents and trace them in the same product.
Langfuse vs LangSmith at a glance
- License
- Langfuse: MIT outside its ee directories. Part of ClickHouse since January 2026. Source 1 Source 2
- LangSmith: Proprietary. The SDKs are MIT. Source
- Self-hosting
- Langfuse: Free, all core features, no usage limits. Docker Compose, Kubernetes, or Terraform. Source 1 Source 2
- LangSmith: An add-on to the Enterprise plan, on Kubernetes with Helm. Hybrid and BYOC also on Enterprise. Source 1 Source 2
- SDKs
- Langfuse: Python and JS/TS, built on OpenTelemetry. Source
- LangSmith: Python, TypeScript, Go, and Java. Source
- Conversations
- Langfuse: Sessions group traces that share a session ID. Source
- LangSmith: Threads group traces by a session_id or thread_id metadata key. Source
- OpenTelemetry
- Langfuse: OTLP over HTTP (JSON or protobuf) at /api/public/otel. No gRPC. Source
- LangSmith: OTLP at your API URL plus /otel. Maps GenAI, TraceLoop, OpenInference, and Logfire attributes. Source
- Production evals
- Langfuse: LLM-as-a-judge on observations; code evaluators in Python or TypeScript. Sessions are not an evaluator target. Source 1 Source 2
- LangSmith: LLM-as-a-judge and code evaluators on filtered, sampled runs; multi-turn evaluators per completed thread. Source 1 Source 2 Source 3
- Offline evals and CI
- Langfuse: Datasets, experiments, annotation queues, and a GitHub Action that fails a workflow on a regression. Source 1 Source 2 Source 3
- LangSmith: Datasets, experiments, pairwise comparisons, annotation queues, and pytest and Vitest or Jest integrations. Source 1 Source 2 Source 3 Source 4 Source 5
- Recurring failures
- Langfuse: Topics is on the roadmap. The Halo assistant can cluster errors when you ask it to. Source 1 Source 2
- LangSmith: Engine (launched in May 2026) clusters issues by severity, billed per run. Insights on Plus and Enterprise. Source 1 Source 2 Source 3
- Prompt management
- Langfuse: Versions and labels, cached client-side by the SDKs, and a playground. Source 1 Source 2 Source 3
- LangSmith: Commits, tags, and Staging and Production environments, cached by the SDK, with a Playground and Prompt Canvas. Source 1 Source 2 Source 3
- Alerts
- Langfuse: Slack, PagerDuty, GitHub Actions, and webhooks. 2 to 100 per organization on Cloud. Source
- LangSmith: Slack, PagerDuty, Dynatrace, and webhooks. Source
- Free plan
- Langfuse: Hobby: 50k units a month, 30 days of data, 2 users. Source
- LangSmith: Developer: 1 seat, 5k base traces a month, kept 14 days. Source
- First paid plan
- Langfuse: Core, $29 a month: 100k units, 90 days, unlimited users. Then $8 per 100k units. Source
- LangSmith: Plus, $39 per seat a month: 10k base traces. Then $5.00 per 1,000 base traces. Source
Tracing and data model
Langfuse's current SDKs, Python v5 and JS/TS v6, are built on OpenTelemetry. You wrap functions with the observe decorator or swap in Langfuse's drop-in OpenAI client by changing one import. Data lands as observations, grouped into traces, and traces that share a session ID form a session.
LangSmith starts from environment variables and the traceable decorator, or wrappers such as wrap_openai and wrap_anthropic. Every unit of work is a run, the runs for one operation form a trace (up to 25,000 runs each), and traces that share a session_id or thread_id metadata key form a thread. Newer LangSmith Cloud workspaces route traces to an agent and environment rather than a project.
One date matters for Langfuse Cloud users: it becomes v4-only on November 16, 2026. Python SDK v2 and JS/TS SDK v3 and older send through the legacy batch ingestion API, which is removed that day, and trace-level evaluators stop producing results. Langfuse says most projects need no migration.
Verdict: Close. Both trace a Python or TypeScript app in a few lines. LangSmith adds Go and Java. Langfuse Cloud is v4-only from November 16, 2026, so older Langfuse SDKs need an upgrade.
Sources: Langfuse docs: SDK overview; Langfuse docs: Instrumentation; Langfuse docs: Data model; Langfuse docs: Sessions; Langfuse changelog: Langfuse v4 (2026-08-17); Langfuse docs: Compatibility; LangSmith docs: Tracing quickstart; LangSmith docs: Observability concepts; LangSmith docs: Threads
Framework lock-in
Neither product requires LangChain. LangSmith's docs say manual instrumentation adds tracing to any code, regardless of the framework, and its integrations page lists the OpenAI Agents SDK, the Vercel AI SDK (JS/TS only), CrewAI, PydanticAI, the Claude Agent SDK, Google ADK, and Mastra next to LangChain and LangGraph. Langfuse lists more than 100 integrations, including LangChain and LangGraph, LlamaIndex, the OpenAI Agents SDK, the Vercel AI SDK, CrewAI, Pydantic AI, and the Claude Agent SDK.
The tie shows up around the tracer. LangSmith Engine's proposed code changes are for agents built with Deep Agents, LangChain, and LangGraph. Getting data out also differs: since August 3, 2026, new LangSmith customers get bulk export on Enterprise only. Langfuse's scheduled export to object storage is free when you self-host, and the code itself is MIT.
Verdict: Langfuse. Both trace any framework, but Langfuse's license, free self-hosting, and export options make it easier to leave.
Sources: LangSmith docs: Observability concepts; LangSmith docs: Integrations; Langfuse integrations; LangSmith docs: Engine; LangSmith docs: Bulk data export; Langfuse docs: Export to blob storage; Langfuse LICENSE on GitHub
OpenTelemetry support
Langfuse receives OTLP at /api/public/otel over HTTP, with JSON or protobuf bodies; gRPC is not supported yet. It aims to follow the OpenTelemetry GenAI semantic conventions and also maps OpenInference, llm.*, MLflow, and langfuse.* attributes. Its own SDKs are a thin layer on the official OpenTelemetry client.
LangSmith receives OTLP at your LangSmith API URL plus /otel, authenticated with an x-api-key header and a project or agent header. It maps GenAI, TraceLoop, OpenInference, and Logfire attributes onto its run fields. The LangSmith SDK can also send its traces to another OpenTelemetry endpoint, and one gotcha is documented: a child span whose parent is never exported does not appear.
Either way, an OpenTelemetry Collector with two exporters can feed both products, or a third one, from the same instrumentation while you compare.
Verdict: A tie on ingest. Langfuse is OpenTelemetry underneath; LangSmith speaks it fluently when you ask.
Sources: Langfuse docs: OpenTelemetry; LangSmith docs: Trace with OpenTelemetry
Evals
The offline loop is well covered on both sides: datasets, experiments, annotation queues, LLM-as-a-judge, and code evaluators. Langfuse ships a GitHub Action that fails a workflow on a regression and documents pytest and Vitest patterns. LangSmith adds pairwise experiments and pairwise annotation queues, and its pytest plugin and Vitest and Jest integrations define evals as test cases.
Production evals split on conversations. Langfuse's LLM-as-a-judge evaluators run on observations or experiments; trace-level evaluators are deprecated, and a session ID works only as a filter. Session, multi-span, and trajectory evaluations are on Langfuse's roadmap. LangSmith's multi-turn evaluators run once per completed thread, after an idle period that defaults to 10 minutes. Its online evaluators run on filtered, sampled runs, and any trace they touch is upgraded to extended retention, at $2.50 per 1,000 traces.
For finding recurring failures, LangSmith Engine, launched in public beta in May 2026, scans traces on a dynamic schedule, clusters what it finds, prioritizes issues by severity, and proposes fixes. It is billed in LangChain Standard Units at $1 each, about 7 to 45 per run, with a default cap of 750 a month. Langfuse's equivalent, Topics, is on its roadmap; today its Halo assistant can cluster errors when you ask it to.
Verdict: LangSmith, for agents that hold conversations: it scores whole threads, and Engine looks for recurring issues today. For offline evals the two are even.
Sources: Langfuse docs: Evaluation overview; Langfuse docs: Annotation queues; Langfuse docs: Experiments in CI/CD; Langfuse docs: LLM-as-a-judge; Langfuse docs: Code evaluators; Langfuse roadmap; Langfuse docs: Halo; LangSmith docs: Evaluation concepts; LangSmith docs: Pairwise evaluations; LangSmith docs: Annotation queues; LangSmith docs: pytest integration; LangSmith docs: Vitest and Jest integration; LangSmith docs: Multi-turn online evaluators; LangSmith docs: LLM-as-a-judge online evaluators; LangSmith docs: Engine; LangChain blog: Introducing LangSmith Engine (May 2026); LangSmith pricing
Prompt management
Langfuse versions prompts and points labels at versions, so a label can stand for an environment, a tenant, or an experiment. The SDKs cache prompts client-side, which Langfuse says removes the latency after first use, and the playground compares variants side by side. Protected labels need the Teams add-on or Enterprise.
LangSmith saves every prompt edit as a commit with a hash, with tags and Staging and Production environments you assign to commits and roll back. The SDK pulls prompts at runtime with caching on by default. The Playground and Prompt Canvas help edit prompts, and Context Hub, added in May 2026, versions bundles of instructions and tools.
Verdict: Even. Both serve versioned prompts to your app at runtime with SDK caching. Decide on the rest of the stack.
Sources: Langfuse docs: Prompt version control; Langfuse docs: Prompt caching; Langfuse docs: Playground; LangSmith docs: Prompt engineering concepts; LangSmith docs: Manage prompts; LangSmith docs: Manage prompts programmatically
Self-hosting and licensing
Langfuse is MIT licensed outside its ee directories, and its self-hosting page offers every core feature for free without limits, including SSO, organization-level RBAC, alerts, and scheduled export. Enterprise adds project-level RBAC, SCIM, audit logs, and data retention controls. You run two containers on PostgreSQL, Redis or Valkey, ClickHouse, and S3-compatible storage, with Docker Compose on one VM, Kubernetes with Helm, or Terraform for AWS, Azure, and GCP. ClickHouse acquired Langfuse in January 2026 and says Langfuse stays open source and self-hostable.
LangSmith is proprietary; its own FAQ answers the open-source question with a no. Self-hosted LangSmith is an add-on to the Enterprise plan, installed on Kubernetes with Helm and a license key. Hybrid and bring-your-own-cloud deployments, where LangChain runs the control plane and your cloud holds the data, are Enterprise too.
Verdict: Langfuse, clearly. If trace data has to stay in your infrastructure and there is no Enterprise budget, it is the only one of the two you can run.
Sources: Langfuse LICENSE on GitHub; Langfuse self-hosted pricing; Langfuse docs: Self-hosting; Langfuse blog: Langfuse joins ClickHouse (2026-01-16); LangSmith docs: FAQ; LangSmith docs: Self-hosted LangSmith; LangSmith docs: Bring your own cloud
Pricing
Langfuse bills units, and a unit is any tracing data point: a trace, an observation, or a score. Hobby is free with 50k units a month, 30 days of data, and 2 users. Core is $29 a month with 100k units, 90 days, and unlimited users; Pro is $199 with 3 years of data, plus $300 for the Teams add-on; Enterprise is $2,499. Past the included units, overage is $8 per 100k up to 1M, then $7 to 10M, $6.50 to 50M, and $6 beyond.
LangSmith bills seats and traces. Developer is free for one seat with 5k base traces a month; Plus is $39 per seat with 10k. After that a base trace costs 0.005 LangChain Standard Units and the extended-retention upgrade 0.0025, at $1 per unit: $5.00 per 1,000 traces kept 14 days, and $2.50 per 1,000 more to keep them 180 days. Engine, Fleet, sandboxes, and Agent Server are metered in the same units.
A worked example, with our assumptions stated. Take 100,000 traces a month, three people, and 10 Langfuse units per trace for the trace, its observations, and its scores. LangSmith Plus is 3 seats at $39, or $117, plus 90,000 traces past the included 10,000 at $5.00 per 1,000, or $450: $567 a month with 14-day retention. Langfuse Core is $29 plus 900,000 units past the included 100,000 at $8 per 100,000, or $72: $101 a month with 90 days of data. At 30 units per trace, Langfuse Core rises to $241. LangSmith stays at $567, and any trace an online evaluator touches adds $2.50 per 1,000 for extended retention.
Verdict: Langfuse is cheaper in both cases above, and the gap grows with every seat. Count your own spans per trace before trusting any example, this one included.
Sources: Langfuse Cloud pricing; LangSmith pricing; LangSmith docs: Usage and billing; LangSmith docs: LLM-as-a-judge online evaluators
Where Moda fits
When to consider Moda instead
Both tools record what your agent did and score what you tell them to score. If the hard part of your week is finding which production failures matter, that is a different job, and it is the one Moda is built for. We make Moda, so weigh this section accordingly.
Moda reads every production trace end to end and detects the failures that never throw an error: frustrated users, failing tools, lazy answers, and hallucinations. Add your own checks in plain language and get alerts in Slack, Linear, or email.
Failures are grouped into problems ranked by how many users they affect. Each problem shows the traces behind it, the turn where it went wrong, and related problems that share the cause.
When your team changes a prompt or model, Moda replays the change against the traces that failed: a simulated user plays each conversation, every tool call is simulated from the tool calls recorded in those traces, and a judge compares pass rates with the current configuration. Your team reviews the result and decides what ships, and Moda can post a Moda evals check on your pull requests. Moda imports Langfuse history; it does not import LangSmith history.
Moda fits when
- Your agents are in production, and nobody has time to read traces or write an evaluator for each failure mode.
- You want failures grouped into problems and ranked by how many users they affect.
- You want to test a prompt or model change against the production traces that failed, without writing tool mocks.
- You already send OpenTelemetry GenAI, OpenLLMetry, or Vercel AI SDK spans, so trying Moda is an exporter change.
Stay with Langfuse or LangSmith when
- You need open-source code or free self-hosting. Moda is neither; Langfuse is both.
- You want to start today on a public price. Moda is invite-only with custom pricing, and both of these have free plans.
- You serve prompts to your app at runtime. Moda versions prompts from your repository, and its SDKs do not fetch them.
- You want evals written as test cases in your own suite, over datasets you curate. Moda builds its eval sets from production traces.
- You need Go or Java SDKs, or a place to deploy agents.
Sources: Moda docs: Signals; Moda docs: Problems; Moda docs: Prompt management; Moda docs: OpenTelemetry; Moda blog: Migrating from Langfuse
Side by side: Moda vs Langfuse · Moda vs LangSmith
Frequently asked questions
Is Langfuse better than LangSmith?
Neither wins everywhere. Langfuse leads on open source, free self-hosting, and price in the cases we worked through. LangSmith leads on LangChain and LangGraph depth, Go and Java SDKs, evaluators that score whole conversations, and Engine for recurring issues. For prompt management and offline evals the two are close.
Is Langfuse open source? Is LangSmith?
Langfuse is MIT licensed outside its ee directories, and its core features can be self-hosted for free without usage limits. LangSmith is proprietary: its SDKs are MIT, and self-hosting is an add-on to the Enterprise plan.
Does LangSmith require LangChain?
No. LangSmith traces any code through its SDKs or OpenTelemetry, and lists integrations for the OpenAI Agents SDK, the Vercel AI SDK, CrewAI, PydanticAI, and others. Some features lean on LangChain's own frameworks: Engine's proposed code changes are for Deep Agents, LangChain, and LangGraph agents.
Can Langfuse trace LangChain and LangGraph apps?
Yes. Langfuse documents a LangChain integration that covers LangGraph, alongside more than 100 other providers, frameworks, and tools.
Which is cheaper, Langfuse or LangSmith?
Langfuse, in the cases we priced: LangSmith charges per seat and per trace, while Langfuse's paid plans include unlimited users. At 100,000 traces a month, three people, and 10 Langfuse units per trace, LangSmith Plus comes to $567 a month and Langfuse Core to $101. Langfuse's bill grows with spans per trace, so price it with your own numbers.
Do Langfuse and LangSmith support OpenTelemetry?
Both accept OTLP over HTTP. Langfuse's endpoint is /api/public/otel and its SDKs are built on OpenTelemetry. LangSmith's is your API URL plus /otel, with mappings for GenAI, TraceLoop, OpenInference, and Logfire attributes. A Collector with two exporters can send the same spans to both.
What changes when Langfuse Cloud moves to v4 on November 16, 2026?
Langfuse Cloud becomes v4-only. Python SDK v2 and JS/TS SDK v3 and older, which send through the legacy batch ingestion API, stop working there, and trace-level evaluators stop producing results. Langfuse says most projects need no migration; check its compatibility page for your SDK versions.
Who owns Langfuse now?
ClickHouse acquired Langfuse in January 2026. Langfuse says it stays open source and self-hostable, with no planned changes to licensing.
Where does Moda fit?
Moda is production monitoring and evals for AI agents. It runs detections on every trace, groups failures into problems ranked by users affected, and replays a prompt or model change against the traces that failed. It is not open source, has no free plan, and does not serve prompts at runtime, so it does not replace either tool for those jobs.
Sources and corrections
Every claim about Langfuse and LangSmith on this page was checked against the vendor’s own page below on the date shown. Both products ship often; if something here is wrong or out of date, report an error and we correct it.
- Langfuse docs: SDK overview, checked November 12, 2026
- Langfuse docs: Instrumentation, checked November 12, 2026
- Langfuse docs: Data model, checked November 12, 2026
- Langfuse docs: Sessions, checked November 12, 2026
- Langfuse changelog: Langfuse v4 (2026-08-17), checked November 12, 2026
- Langfuse docs: Compatibility, checked November 12, 2026
- Langfuse integrations, checked November 12, 2026
- Langfuse docs: OpenTelemetry, checked November 12, 2026
- Langfuse docs: LLM-as-a-judge, checked November 12, 2026
- Langfuse docs: Code evaluators, checked November 12, 2026
- Langfuse docs: Evaluation overview, checked November 12, 2026
- Langfuse docs: Annotation queues, checked November 12, 2026
- Langfuse docs: Experiments in CI/CD, checked November 12, 2026
- Langfuse roadmap, checked November 12, 2026
- Langfuse docs: Halo, checked November 12, 2026
- Langfuse docs: Prompt version control, checked November 12, 2026
- Langfuse docs: Prompt caching, checked November 12, 2026
- Langfuse docs: Playground, checked November 12, 2026
- Langfuse LICENSE on GitHub, checked November 12, 2026
- Langfuse self-hosted pricing, checked November 12, 2026
- Langfuse docs: Self-hosting, checked November 12, 2026
- Langfuse blog: Langfuse joins ClickHouse (2026-01-16), checked November 12, 2026
- Langfuse Cloud pricing, checked November 12, 2026
- Langfuse docs: Alerts, checked November 12, 2026
- Langfuse security and compliance, checked November 12, 2026
- Langfuse HIPAA, checked November 12, 2026
- Langfuse docs: Export to blob storage, checked November 12, 2026
- LangSmith docs: Tracing quickstart, checked November 12, 2026
- LangSmith docs: Observability concepts, checked November 12, 2026
- LangSmith docs: Log traces to an agent (SDK versions by language), checked November 12, 2026
- LangSmith docs: Threads, checked November 12, 2026
- LangSmith docs: Integrations, checked November 12, 2026
- LangSmith docs: Trace with OpenTelemetry, checked November 12, 2026
- LangSmith docs: Evaluation concepts, checked November 12, 2026
- LangSmith docs: Pairwise evaluations, checked November 12, 2026
- LangSmith docs: LLM-as-a-judge online evaluators, checked November 12, 2026
- LangSmith docs: Code online evaluators, checked November 12, 2026
- LangSmith docs: Multi-turn online evaluators, checked November 12, 2026
- LangSmith docs: Annotation queues, checked November 12, 2026
- LangSmith docs: pytest integration, checked November 12, 2026
- LangSmith docs: Vitest and Jest integration, checked November 12, 2026
- LangSmith docs: Insights, checked November 12, 2026
- LangSmith docs: Engine, checked November 12, 2026
- LangChain blog: Introducing LangSmith Engine (May 2026), checked November 12, 2026
- LangSmith docs: Prompt engineering concepts, checked November 12, 2026
- LangSmith docs: Manage prompts, checked November 12, 2026
- LangSmith docs: Manage prompts programmatically, checked November 12, 2026
- LangSmith docs: FAQ, checked November 12, 2026
- LangSmith docs: Self-hosted LangSmith, checked November 12, 2026
- LangSmith docs: Bring your own cloud, checked November 12, 2026
- LangSmith docs: Alerts, checked November 12, 2026
- LangChain Trust Center, checked November 12, 2026
- LangSmith pricing, checked November 12, 2026
- LangSmith docs: Usage and billing, checked November 12, 2026
- LangSmith docs: Export traces, checked November 12, 2026
- LangSmith docs: Bulk data export, checked November 12, 2026
- Moda docs: Signals, checked November 12, 2026
- Moda docs: Problems, checked November 12, 2026
- Moda docs: Prompt management, checked November 12, 2026
- Moda docs: OpenTelemetry, checked November 12, 2026
- Moda blog: Migrating from Langfuse, checked November 12, 2026