Langfuse vs LangSmith: Features, Pricing, and Which to Pick

If you're building an LLM app or agent in 2026, you've probably narrowed observability down to these two. Langfuse and LangSmith both trace, evaluate, and manage prompts, and their feature lists overlap heavily.

They differ in four places: who owns the code, where it can run, how tightly it fits LangChain, and how the bill grows as your traffic does. Both tools also shipped major changes in 2026.

We checked both products' pricing pages, docs, and changelogs on October 22, 2026, and laid out where each one wins, so you can pick the right one for your stack in one read.

Quick Verdict: Langfuse vs LangSmith

Pick Langfuse if you want an MIT-licensed core you can self-host for free, framework-neutral OpenTelemetry tracing, and pricing with no seat fees.

Pick LangSmith if you build on LangChain or LangGraph and want native tracing plus managed agent deployment from the same company that makes your framework.

Langfuse vs LangSmith at a Glance

Here are the main differences to be aware of:

What Is Langfuse?

Langfuse is an open-source platform to "trace, evaluate, and improve AI agents." It captures traces of LLM calls, tools, and retrieval, then layers on evals, prompt management, datasets, dashboards, and alerts.

ClickHouse acquired Langfuse in January 2026 and says there are no planned licensing changes. Langfuse v4, released in August 2026, added monitors and alerts, code evaluators, full-text search, and much faster queries and dashboards. Langfuse reports 50,000+ companies using it, including Canva.

What Is LangSmith?

LangSmith is LangChain’s “framework-agnostic platform for testing, deploying, monitoring, and improving your agents”. It's the observability and evaluation layer for LangChain and LangGraph, and it also runs agents through LangSmith Deployment.

In 2026, LangSmith moved usage billing to LangChain Standard Units ($1 each) and added Engine, which clusters production failures into issues and proposes fixes as pull requests. Customers include Klarna and Madrigal Pharmaceuticals.

Open Source and Self-Hosting

For many teams, this section settles the choice.

Langfuse's core is MIT licensed. You can self-host every core feature for free with no usage limits. Enterprise extras like SCIM, audit logs, and data retention policies sit in a separately licensed folder and need a license key.

The catch is that a production Langfuse deployment runs two app containers plus Postgres, ClickHouse, Redis, and S3-compatible storage. That's a big commitment.

LangSmith is closed source. Self-hosting is an add-on to the Enterprise plan, designed for LangChain's largest customers, and also runs on Kubernetes with ClickHouse, Postgres, Redis, and blob storage. BYOC, where LangChain runs LangSmith in your own cloud account, is the middle option.

If data must stay in your environment and you don't have an enterprise budget, Langfuse is the only one of the two that fits.

Langfuse vs LangSmith Pricing

Both tools have a free tier and usage-based overage, but they meter different things.

Langfuse Pricing

Langfuse Cloud has no seat fees on any plan:

• Hobby: free, 50,000 units a month, 30 days of data, 2 users.

• Core: $29 a month, 100,000 units, 90 days of data, unlimited users.

• Pro: $199 a month, 3 years of data, SOC 2 and ISO 27001 reports, and a HIPAA-ready region.

• Enterprise: $2,499 a month, with audit logs, SCIM, and SLAs.

A unit is any data point you send, such as a trace, each observation inside it (spans, generations, events), and each score. Overage is $8 per 100,000 units and drops at volume. Our Langfuse pricing breakdown has the full table.

LangSmith Pricing

LangSmith charges per seat plus usage:

• Developer: free for 1 seat, up to 5,000 base traces a month.

• Plus: $39 per seat a month, 10,000 base traces included, unlimited seats.

• Enterprise: custom, with self-hosted and hybrid options.

Beyond the included traces, the pricing page lists 0.005 LSU per base trace and 0.0025 LSU to upgrade a trace to extended retention, at $1 per LSU. Base traces keep data for 14 days, and extended traces keep it for 180 days.

What That Looks Like at Real Volume

Say a five-person team sends 50,000 traces a month, each with about 10 spans. By our math at October 2026 list prices:

• Langfuse Core: about 550,000 units, so $29 plus roughly $36 in overage, around $65 a month before eval scores, which also count as units.

• LangSmith Plus: five seats at $195 plus 40,000 traces beyond the included 10,000 at $0.005 each, around $395 a month, or about $520 if you keep every trace past 14 days.

Your numbers will move with span counts, retention needs, and seats, so plug in your own traffic. The gap mostly comes from LangSmith's seat fees and per-trace pricing.

Framework Fit: LangChain-Native vs Framework-Neutral

LangSmith is the lowest-friction choice for LangChain and LangGraph. Tracing switches on with an environment variable, agent trajectories show up with the right structure, and the same company ships the framework and the observability.

It also works outside LangChain, with SDKs for Python, TypeScript, Go, and Java, and support for the OpenAI SDK, Anthropic SDK, Vercel AI SDK, CrewAI, and OpenTelemetry. You give up some of the zero-config setup.

Langfuse is framework-neutral and built on OpenTelemetry. It has native SDKs for Python and TypeScript, OTel support for other languages, and 100+ integrations, including a LangChain and LangGraph callback handler.

If your stack mixes frameworks, or you might switch, that neutrality is worth a lot.

Both accept OpenTelemetry data, and OpenTelemetry's GenAI semantic conventions give LLM spans a shared format, so moving instrumentation between them takes less work.

Evaluations, Prompts, and Alerting

On core features, the two are close.

Evaluations

Both support datasets, experiments, LLM-as-judge and code evaluators, human annotation queues, and evals in CI.

LangSmith adds multi-turn evaluators on production traces and judge calibration, and Langfuse added code evaluators in v4. For how to structure evals in either tool, see our guide to AI agent evals.

Prompt Management

Langfuse offers prompt management with versioning, release labels, composable prompts, caching, and a playground. LangSmith has its Prompt Hub (now called Prompt & Context Hub in the docs) and a playground.

Alerting and Monitoring

Langfuse now has threshold alerts routed to Slack, webhooks, and GitHub Actions, with 2 to 100 alerts per organization depending on plan and no limit when self-hosted.

LangSmith offers online evals, dashboards for latency and error rates, and alerts routed to Slack, PagerDuty, Dynatrace, or any webhook.

If older comparisons told you Langfuse had no alerts, that changed with v4.

Agents and Deployment

LangSmith covers more of the agent lifecycle. LangSmith Deployment is a managed runtime for LangGraph and other agent frameworks, with MCP and A2A support, and Engine analyzes production failures and suggests fixes. Fleet adds no-code agents, and an LLM Gateway is in the product line too.

Langfuse focuses on observability, evals, and prompts. It doesn't host or run your agents, so you deploy them wherever you like and send traces to Langfuse.

Langfuse Pros and Cons

Pros:

✅ MIT-licensed core you can self-host for free, with no usage limits

✅ No seat fees on any cloud plan

✅ Framework-neutral and OpenTelemetry-based

✅ SOC 2 Type II and ISO 27001 audits, GDPR, and a HIPAA-ready region on Pro

Cons:

❌ Self-hosting in production means running Postgres, ClickHouse, Redis, and object storage

❌ No managed agent runtime

❌ Self-hosted SCIM, audit logs, and retention policies need an enterprise license key

LangSmith Pros and Cons

Pros:

✅ Deepest integration with LangChain and LangGraph

✅ Managed agent deployment and Engine in the same platform

✅ SOC 2 Type II certified, and GDPR and HIPAA compliant

✅ Multi-turn evaluators on production traces, plus judge calibration with human feedback

Cons:

❌ Closed source, with self-hosting only as an Enterprise add-on

❌ Per-seat pricing plus per-trace usage adds up for bigger teams

❌ The free plan has one seat, and base traces expire after 14 days

Which Should You Choose?

Choose Langfuse if:

• You need self-hosting without an enterprise contract.

• Your stack mixes frameworks, or you want to stay framework-neutral.

• You have a larger team and want to avoid seat fees.

Choose LangSmith if:

• You build on LangChain or LangGraph and want zero-config tracing.

• You want managed agent deployment from the same vendor.

• You want multi-turn evaluators and judge calibration on production traces.

If neither fits, our roundup of Langfuse alternatives covers the wider field.

Migrating From LangSmith to Langfuse (or Back)

Langfuse publishes a migration guide from LangSmith. Langfuse Cloud becomes v4-only on November 16, 2026, so point new integrations at the v4 SDKs.

The short version is to keep your traceable code and send the same spans to Langfuse through LangSmith's OpenTelemetry mode, or add the Langfuse callback handler for LangChain apps, then copy datasets and prompts over with the LangSmith SDK.

Going the other way, LangSmith ingests OpenTelemetry data, so OTel-instrumented apps can point their exporter at LangSmith. Either way, plan to recreate evaluators and re-run experiments, since scores don't carry over cleanly.

What Neither Tool Covers for Conversational Agents

Langfuse and LangSmith both tell you what your agent did. If you're building voice or chat agents, the bigger question is whether the agent handles a real person who interrupts, mumbles, changes their mind, or tries to talk it out of a policy, and you want that answer before launch.

Neither tool simulates those conversations natively. Both point to open-source libraries like OpenEvals for simulated multi-turn users, and Langfuse adds partner integrations for voice. Cekura is built for that layer, and it runs alongside either one.

The failure rate is higher than most teams expect. In Cekura’s public voice benchmarks, the most reliable of nine voice agent platforms passed 87.8% of 82 caller scenarios on all three runs, and the least reliable passed 30.5%.

Across 12 realtime speech-to-speech models and 3,198 live calls, scores ranged from 81.7% down to 41.5%. Simulation is how you find out which of those scenarios your own agent fails before a customer finds them.

Native integrations work out of the box for Retell, VAPI, ElevenLabs, LiveKit, Pipecat, Bland, and more.

With Cekura, teams can:

• Simulate thousands of conversations with personas, accents, background noise, and interruptions before go-live.

• Test voice agents end to end, including latency, turn-taking, and speech quality.

• Score every run with 33 predefined metrics and custom Python metrics.

• Gate releases in CI/CD with Tests as Code on every prompt or model change.

• Replay real production calls with the original captured audio to confirm fixes.

• Ingest OpenTelemetry traces and monitor production conversations with alerts.

It’s SOC 2-, HIPAA-, and GDPR-compliant for transcript redaction, role-based access, and audit trails.

Book a demo with Cekura to see how it tests voice and chat agents alongside Langfuse or LangSmith.

Frequently Asked Questions

Are Langfuse and LangSmith the same?

No, Langfuse and LangSmith are separate products from different companies that cover much of the same ground. Both trace LLM calls, run evals, and manage prompts.

Langfuse has an MIT-licensed core you can self-host for free and charges no seat fees. LangSmith is closed source, charges per seat, and adds managed agent deployment.

Is Langfuse part of LangChain?

No, Langfuse isn’t part of LangChain. ClickHouse acquired Langfuse in January 2026, and it connects to LangChain and LangGraph through a callback handler.

Can LangSmith be self-hosted or run locally?

Yes, but self-hosting LangSmith is an add-on to the Enterprise plan. On any plan, you can send traces from local development to LangSmith’s cloud.

If you want data kept in your own environment without running it yourself, LangChain also offers BYOC, where it runs LangSmith in your cloud account. Langfuse, by contrast, can run locally or on your own servers for free under its MIT core.

Is LangSmith free or paid?

LangSmith has a free Developer plan for one seat with 5,000 base traces a month. The Plus plan costs $39 per seat a month with 10,000 base traces included, and extra traces cost $5 per 1,000. Enterprise is custom.

Does LangSmith work without LangChain?

Yes, LangSmith works without LangChain. It traces apps built with the OpenAI SDK, Anthropic SDK, Vercel AI SDK, CrewAI, custom code, and OpenTelemetry, though setup takes more work than with LangChain.

What changed for Langfuse after the ClickHouse acquisition?

Langfuse says it stays open source and self-hostable with no planned licensing changes. Langfuse Cloud kept running as before, and self-hosted Enterprise is now sold bundled with ClickHouse Cloud, BYOC, or Private.

Is Langfuse or LangSmith better for HIPAA workloads?

Both can handle HIPAA workloads, on different terms. Langfuse offers a HIPAA-ready region with a BAA on Pro and above, and you can self-host it inside your own environment.

LangSmith says it is SOC 2 Type II, ISO 27001, GDPR, and HIPAA compliant, so ask LangChain about BAA terms for your plan.

Can Langfuse or LangSmith test voice agents?

Both can trace and score voice agent calls, but neither runs end-to-end voice simulations natively. Both point to open-source libraries like OpenEvals for simulated users. A dedicated testing layer such as Cekura runs those calls before launch and sends results alongside your traces.