LLM applications fail differently than traditional software. A prompt change can quietly degrade output quality for days before anyone notices. A RAG pipeline might retrieve the right documents but synthesize them poorly. An agent loop can burn through thousands of tokens without accomplishing its task. Without observability, these failures surface as vague user complaints long after the damage is done.
Langfuse and LangSmith are two platforms built to solve this problem. Both provide tracing, evaluation, and prompt management for LLM applications. But they differ fundamentally in architecture, pricing transparency, and deployment philosophy. Langfuse is open source and framework agnostic, designed for teams that want full control over their observability stack. LangSmith is a managed platform from the LangChain team, optimized for teams already deep in the LangChain ecosystem.
This guide compares both platforms across tracing depth, evaluation workflows, deployment models, and total cost of ownership. Each section includes real cost scenarios, sourced limitations, and specific use cases where one platform clearly leads.
Quick Comparison: Langfuse vs LangSmith
| Langfuse | LangSmith | |
|---|---|---|
| Best for | Open source self hosting, framework agnostic teams, data sovereignty | LangChain/LangGraph users, managed SaaS, teams prioritizing ecosystem integration |
| Pricing model | Free self host (MIT license) · Cloud Hobby free 50k obs/month · Core $29/mo 100k obs · Pro $199/mo unlimited | Free 5k traces/month · Plus $39/seat/month 10k traces · Enterprise custom |
| Architecture | Open source (MIT), runs on ClickHouse, self hostable | Proprietary SaaS, BYOC and self hosted require Enterprise license |
| OpenTelemetry | Native OTLP ingestion, full OTel compatibility | Full OTel support as of March 2026 |
| Framework support | Framework agnostic: LangChain, LlamaIndex, OpenAI SDK, custom stacks | Best with LangChain/LangGraph, supports other frameworks via @traceable decorator |
| Evaluation | LLM-as-judge, custom scoring, annotation workflows | Systematic dataset-based evaluation, built-in evaluators, human annotation queues |
| Deployment | Self hosted (Docker, Kubernetes) · Cloud managed | Cloud SaaS · BYOC (Enterprise) · Self hosted (Enterprise, Kubernetes only) |
| Data residency | Full on prem option, all data stays local | Cloud default, BYOC/self hosted require Enterprise license |
| Alerting | Via Metrics API and integrations, not native in UI | Native configurable alerts with webhooks |
Langfuse Overview
Langfuse is an open source LLM observability platform built for teams that want framework agnostic tracing, transparent pricing, and the option to self host. In January 2026, ClickHouse acquired Langfuse as part of a $400M Series D round, signaling the strategic importance of LLM observability in modern data infrastructure.
Langfuse captures traces as structured observations: spans for work units, generations for LLM calls, and events for point-in-time actions. Its data model is explicit and nested, making complex agent workflows tractable. Because it is built on OpenTelemetry, Langfuse integrates with any stack and can stitch together distributed traces across microservices.
The platform provides evaluation through LLM-as-judge, custom scoring functions, and annotation workflows. Prompt management includes versioning, a playground for iteration, and flexible dashboards that support multi-level aggregations across tracing data. The core platform is MIT licensed and runs on Docker Compose for development or Kubernetes for production. Post acquisition, Langfuse remains fully open source with no planned licensing changes.
Langfuse has grown to 23,000 GitHub stars, with adoption at 19 of the Fortune 50. A medium scale self hosted deployment costs approximately $3,000 to $4,000 per month in infrastructure, compared to $199 to $300 per month for the equivalent cloud Pro tier at mid market scale.
Strengths: Full data ownership, transparent pricing, framework agnostic design, self hosted option with no license restrictions, explicit trace schema that simplifies debugging complex agent flows.
Limitations: Evaluation framework is less opinionated than LangSmith’s — you get scoring primitives and annotation queues, but need to assemble your own evaluation pipelines. Self hosting requires operational investment. The ecosystem of pre-built integrations and tutorials is smaller than LangSmith’s.
LangSmith Overview
LangSmith is the managed observability platform from the team behind LangChain. It started as a tightly coupled companion to LangChain but has since expanded to support OpenAI SDK, Anthropic SDK, Vercel AI SDK, LlamaIndex, and custom implementations through the @traceable decorator or OpenTelemetry integration.
LangSmith uses the Run Tree model, where each step in a sequence becomes a run with parent/child relationships. A parent run might be an entire agent execution, with child runs for tool calls, prompt formatting, LLM calls, and output parsing. Because it integrates directly with LangChain, tracing is often automatic. The UI allows you to expand any trace to view the details of each node, including streaming tokens.
The evaluation framework is opinionated and systematic. You can build datasets directly from production traces, define custom evaluators, and run experiments comparing prompt changes or model swaps. The playground lets you iterate on prompts against real production data, shortening the feedback loop. LangSmith also includes LangGraph Studio, a specialized IDE for visualizing and debugging agentic applications with breakpoints, state modification, and real time token usage tracking.
LangSmith offers managed cloud, bring-your-own-cloud (BYOC), and self hosted options. The free Developer tier includes 5,000 traces per month. The Plus plan runs $39 per seat per month with 10,000 included traces. Additional traces cost approximately $2.50 to $5.00 per 1,000 traces depending on retention. Enterprise pricing is custom. Both BYOC and self hosted options require an Enterprise license. The platform itself is proprietary, though the client SDKs are MIT licensed.
Strengths: Deep LangChain integration, mature evaluation framework, native alerting and dashboards, LangGraph Studio for agent debugging, managed SaaS with minimal operational overhead.
Limitations: Proprietary platform with no open source self host option (self hosting requires Enterprise license). Seat-based pricing compounds for larger teams. Evaluation features work best within the LangChain ecosystem. Teams not using LangChain will do more manual integration work.
Tracing: Depth and Structure
Both platforms capture hierarchical traces, but they model them differently and optimize for different workflows.
Langfuse tracing model
Langfuse captures traces as collections of observations. An observation can be a span (a unit of work like a function call or retrieval step), a generation (a specialized span for an LLM call with inputs, outputs, metadata, token counts, and cost), or an event (a point-in-time action like a user click). Observations are nested and typed, which makes complex flows tractable.
Because Langfuse is built on OpenTelemetry, it integrates with any framework and can stitch together distributed traces across microservices. You get a detailed, nested view of your application where a parent agent span leads to a child tool span and then an LLM generation. It tracks cost, latency, and tokens for each step.
Langfuse recently released specific features to make it radically easier to understand and debug complex, multi step agents, which directly addresses the challenge of tracing agentic workflows that often span multiple LLM calls and tool invocations.
LangSmith tracing model
LangSmith uses the Run Tree model, native to the LangChain Expression Language (LCEL). Each step in a sequence becomes a run with parent/child relationships nested to form a tree. A parent run might be an entire agent execution. Child runs include tool calls, prompt formatting, LLM calls, and output parsing.
Because it integrates with LangChain, tracing is often automatic. The @traceable decorator lets you wrap any custom Python function and trace it as a run. You can trace inputs, outputs, total tokens, costs, start/end time, errors, and streaming tokens. The UI allows you to expand any trace to view the details of each node.
LangSmith also records events like streaming tokens, giving real time visibility into how responses are generated.
Winner: Langfuse. Both systems support hierarchical traces, but Langfuse’s OpenTelemetry foundation and language agnostic SDKs give it an edge in flexibility. Teams running polyglot stacks or distributed systems will find Langfuse easier to integrate without reworking their instrumentation.
Evaluation Workflows
Evaluation is the most critical feature for production LLM applications. Both platforms provide evaluation capabilities, but they differ in opinionation and workflow maturity.
Langfuse evaluation
Langfuse provides evaluation through LLM-as-judge scoring, custom Python evaluators, and human annotation workflows. It supports both batch (offline) and in-production (online) evaluation. You create datasets of input/output examples, run your agent against the dataset, and score performance. Results are tracked over time to detect regressions.
Langfuse also supports continuous online evaluation where production traces are automatically scored using LLM-as-judge or custom functions. This helps catch quality drift before it compounds.
The evaluation framework is less opinionated than LangSmith’s. You get scoring primitives, annotation queues, and LLM-as-judge, but need to assemble your own evaluation pipelines. This gives flexibility but requires more upfront design work.
LangSmith evaluation
LangSmith’s evaluation framework is systematic and opinionated. You can build datasets directly from production traces, which shortens the feedback loop between observing a failure and testing a fix. The platform supports built-in evaluators, custom evaluators, and human annotation queues.
The playground lets you iterate on prompts against real production data, testing changes before deployment. You can run experiments comparing prompt changes, model swaps, or parameter adjustments, with results tracked in a structured experiment log.
LangSmith also supports gold standard evaluations where you define expected outputs and measure how closely your model matches them. The evaluation UI highlights regressions and makes it easy to drill into specific failures.
Winner: LangSmith. The opinionated evaluation framework, tight integration with datasets built from production traces, and structured experiment tracking give LangSmith a clear lead. Teams that want systematic evaluation workflows without building their own pipeline will benefit from LangSmith’s built-in structure.
Deployment Models and Data Residency
Deployment model determines where your telemetry data lives, who manages the infrastructure, and what compliance frameworks you can meet.
Langfuse deployment
Langfuse offers two deployment paths:
Self hosted: The core platform is MIT licensed and runs on Docker Compose for development or Kubernetes for production. You control the infrastructure, data residency, and retention policies. Langfuse runs on ClickHouse as its analytics backend, which makes the ClickHouse acquisition a natural architectural fit. A medium scale self hosted deployment costs approximately $3,000 to $4,000 per month in infrastructure.
Cloud managed: Langfuse offers a managed cloud service. The Hobby tier is free with 50,000 units per month (traces, observations, and scores each count as one unit) and 30 day retention. Paid cloud plans start from there. For context, a medium scale deployment costs $199 to $300 per month for the cloud Pro tier at mid market scale.
Post acquisition, Langfuse remains fully open source with no planned licensing changes. Self hosting is free with no Enterprise license required.
LangSmith deployment
LangSmith offers three deployment models:
Cloud: Fully managed by LangChain, with data stored in GCP us-central-1 or an EU region depending on your plan.
BYOC (Bring Your Own Cloud): SaaS control plane with self hosted data plane. You manage the infrastructure where traces are stored, but LangChain manages the platform itself.
Self hosted: Fully self managed on your Kubernetes cluster (AWS, GCP, or Azure). You control the infrastructure and data residency.
Both BYOC and self hosted options require an Enterprise license. There is no open source self host option. The platform itself is proprietary, though the client SDKs are MIT licensed.
The free Developer tier includes 5,000 traces per month with 14 day retention. The Plus plan runs $39 per seat per month with 10,000 included traces. Additional traces cost approximately $2.50 to $5.00 per 1,000 traces depending on retention. Enterprise pricing is custom.
Winner: Langfuse. The MIT licensed self host option with no Enterprise license requirement gives Langfuse a clear lead for teams that want full data ownership without negotiating a custom contract. LangSmith’s self host option is available but restricted to Enterprise customers.
Alerting and Monitoring
Real time alerting determines how quickly you detect issues in production.
Langfuse alerting
Langfuse provides a real time analytics dashboard that monitors cost, latency, and quality scores. You can filter dashboards by user, session, or prompt version. For security monitoring, you can track scores over time and manually inspect traces to investigate issues.
Langfuse currently lacks native trace-level alerting directly from the UI. You can subscribe to trace events via webhooks or use the Metrics API to build custom alerting on top of Langfuse data. This gives flexibility but requires engineering work to set up.
LangSmith alerting
LangSmith has mature out of the box alerting. It provides pre-built dashboards automatically generated for each project, monitoring trace counts, error rates, token usage, costs, and tool performance. You can also build custom dashboards to track specific metrics.
Native alerting lets you configure alerts to trigger on events. For example, you define conditions like error rate above 5% over 5 minutes, and LangSmith sends notifications via Slack, email, or webhooks when they are met. Alerts can hook into LangSmith’s automation engine or webhook integrations.
LangSmith also includes an Insights section for anomaly detection and proactive alerts.
Winner: LangSmith. The native alerting with threshold-based triggers, webhook integrations, and anomaly detection gives LangSmith a clear lead. Langfuse is catching up but currently requires custom tooling to achieve the same level of alerting maturity.
Pricing Comparison
Pricing models differ fundamentally between the two platforms. Langfuse uses observation-based pricing with free self hosting. LangSmith uses seat-based pricing with trace volume charges.
Langfuse pricing
Self hosting: Free with MIT license. Infrastructure costs depend on scale. A medium scale deployment costs approximately $3,000 to $4,000 per month in infrastructure.
Cloud Hobby tier: Free with 50,000 units per month (traces, observations, and scores each count as one unit) and 30 day retention.
Cloud Core tier: $29 per month for 100,000 observations, 90 day retention, and unlimited users.
Cloud Pro tier: $199 per month for unlimited data access, higher limits, and unlimited users.
Cloud Enterprise tier: $2,499 per month with custom retention, custom limits, and dedicated support.
For reference, a team processing 500,000 observations per month (typical for a mid market product with moderate agent usage) would pay $199 per month on Langfuse Cloud Pro, or could self host for approximately $3,000 to $4,000 per month in infrastructure if they want full data control.
Pricing details are available at Langfuse pricing.
LangSmith pricing
Developer tier: Free with 5,000 base traces per month, 14 day retention, and 1 seat.
Plus tier: $39 per seat per month with 10,000 base traces per month, 14 day base retention or 400 day extended retention, and up to 10 seats.
Enterprise tier: Custom pricing with custom trace volumes, custom retention, and unlimited seats.
Trace pricing:
Base traces: $0.50 per 1,000 traces with 14 day retention.
Extended traces: $2.50 per 1,000 traces to upgrade to 400 day retention.
Traces with user feedback automatically upgrade to extended retention.
For reference, a 10 person team processing 500,000 traces per month would pay:
10 seats at $39 per seat = $390 per month.
490,000 additional traces beyond the included 10,000 at $0.50 per 1,000 = $245 per month.
Total: $635 per month for base retention, or approximately $1,615 per month if extended retention is needed for all traces.
Pricing details are available at LangSmith pricing.
Cost comparison scenario
Scenario: Mid market team, 10 engineers, 500,000 traces per month, 30 day retention.
Langfuse Cloud Pro: $199 per month (unlimited users, unlimited observations, 90 day retention).
LangSmith Plus: $390 per month (seats) + $245 per month (traces) = $635 per month for 14 day retention, or $1,615 per month for 400 day retention.
Langfuse self hosted: $3,000 to $4,000 per month in infrastructure (full data control, unlimited retention, no per-trace fees).
Pricing based on publicly available information as of April 2026. Enterprise discounts, custom contracts, and negotiated rates are not reflected here.
Framework Support and Integration
Both platforms support multiple frameworks, but their integration depth varies.
Langfuse
Langfuse is framework agnostic from the ground up. It integrates with LangChain, LlamaIndex, OpenAI SDK, LiteLLM, Vercel AI SDK, and any custom stack via Python and TypeScript SDKs. It also supports OpenTelemetry, meaning you can feed LLM traces into the same pipeline as your existing application telemetry.
The data model is clean: traces contain observations (spans, generations, events, and dedicated types for agents, tools, chains, and retrievers), each with structured metadata. This makes it easy to integrate Langfuse into polyglot stacks without reworking your instrumentation.
LangSmith
LangSmith provides SDKs for Python, TypeScript, Go, and Java. It traces applications built with OpenAI SDK, Anthropic SDK, Vercel AI SDK, LlamaIndex, and custom implementations, not just LangChain. That said, the deepest integration remains with LangChain and LangGraph, where tracing is essentially automatic.
For other frameworks, you wrap functions with the @traceable decorator or use the OpenTelemetry integration. As of March 2026, LangSmith supports full end-to-end OpenTelemetry, making it easier to integrate into existing observability stacks.
Winner: Langfuse. The framework agnostic design and native OpenTelemetry support give Langfuse an edge for teams running polyglot stacks or distributed systems. LangSmith works well with multiple frameworks but is optimized for LangChain/LangGraph users.
Who Should Choose Langfuse
Langfuse is the better choice for:
Teams that want full data ownership: The MIT licensed self host option with no Enterprise license requirement makes Langfuse the clear choice for teams with data residency, HIPAA, or GDPR requirements.
Framework agnostic teams: If you are not deeply invested in LangChain, Langfuse’s framework agnostic design and native OpenTelemetry support will save integration work.
Cost sensitive teams: Transparent observation-based pricing with no seat fees makes costs predictable as teams grow. Self hosting eliminates per-trace fees entirely.
Open source first teams: The open source codebase and active community make Langfuse the default choice for teams that value transparency and the ability to inspect or modify the platform.
Who Should Choose LangSmith
LangSmith is the better choice for:
LangChain/LangGraph users: If your stack is built on LangChain, LangSmith’s automatic tracing, LangGraph Studio, and tight ecosystem integration will save significant setup time.
Teams prioritizing evaluation workflows: The opinionated evaluation framework, dataset-driven experiments, and structured testing workflows give LangSmith a clear lead for teams that want systematic evaluation without building their own pipeline.
Managed SaaS preference: If you want a fully managed platform with minimal operational overhead, LangSmith’s cloud offering with native alerting and pre-built dashboards is easier to adopt than self hosting Langfuse.
Enterprise teams with budget: The seat-based pricing is viable for smaller teams. Enterprise plans include BYOC and self hosted options, though both require an Enterprise license.
Verdict
Both Langfuse and LangSmith are production grade LLM observability platforms, but they serve different priorities.
Choose Langfuse if you want full data ownership, transparent pricing, framework agnostic design, and the option to self host without an Enterprise license. It is the best choice for teams with data residency requirements, cost constraints, or polyglot stacks.
Choose LangSmith if you are deeply invested in LangChain/LangGraph and want a managed platform with mature evaluation workflows, native alerting, and minimal operational overhead. It is the best choice for teams that prioritize ecosystem integration and systematic testing over cost or data sovereignty.
For teams outside the LangChain ecosystem, Langfuse’s framework agnostic design and open source model make it the more flexible platform. For LangChain users, LangSmith’s tight integration and opinionated workflows reduce setup friction significantly.
Disclaimer: The information in this article reflects the latest details available at the time of publication and may change as technologies and products evolve. Features, pricing, and plan limits can change over time. Always verify the latest information directly with the vendor before making purchasing or deployment decisions.
Frequently Asked Questions
Is Langfuse really free to self host?
Yes. Langfuse is MIT licensed and free to self host with no Enterprise license required. You only pay for the infrastructure costs to run it on your own cloud or data center.
Does LangSmith work without LangChain?
Yes. LangSmith supports OpenAI SDK, Anthropic SDK, Vercel AI SDK, LlamaIndex, and custom implementations through the @traceable decorator or OpenTelemetry integration. However, the deepest integration remains with LangChain and LangGraph.
What is the actual cost difference at scale?
For a mid market team with 10 engineers and 500,000 traces per month, Langfuse Cloud Pro costs $199 per month. LangSmith Plus costs $635 per month for base retention or $1,615 per month for extended retention. Self hosting Langfuse costs approximately $3,000 to $4,000 per month in infrastructure.
Can I migrate from LangSmith to Langfuse?
Yes. Both platforms support OpenTelemetry, which makes migration feasible. You will need to rework evaluation pipelines and prompt management workflows, but tracing instrumentation can be migrated incrementally.
Does Langfuse support alerting?
Langfuse provides real time dashboards and a Metrics API for custom alerting, but it does not yet have native trace-level alerting directly from the UI. You can subscribe to trace events via webhooks or build custom alerting on top of Langfuse data.
Which platform has better evaluation workflows?
LangSmith has a more opinionated and mature evaluation framework with dataset-driven experiments, built-in evaluators, and structured testing workflows. Langfuse provides evaluation primitives but requires more upfront design work to assemble custom pipelines.
Does LangSmith have an open source option?
No. LangSmith is proprietary. Self hosting requires an Enterprise license. The client SDKs are MIT licensed, but the platform itself is closed source.