GILD verdict
Choose Langfuse when traces stay on your servers or many people need a flat fee, and LangSmith for LangGraph agents and its release gates.
- The workflow
- an AI team choosing where its agent traces, evals and prompts will live once agents serve users
- The decisive evidence
- Reported4 of 5 named teams adopted one; What teams report below sets out each account.
- DocumentedLangfuse's MIT core self-hosts free with every core feature, while self-hosted LangSmith is an add-on to the Enterprise plan.1624
- GILD analysisGILD's calculation from published rates puts 200,000 traces a month for 10 users, kept at least 30 days, at US$1,840 on LangSmith Plus and US$369 on Langfuse Pro. How many of your traces could live on the 14-day base tier is unresolved, and it decides up to US$500 of that gap.
- The tradeoff
- Self-hosting moves the operations work to your team. GAIA's self-hosted Langfuse collector caused 2,359 of the 3,890 Sentry events it logged over 30 days, which a pull request opened on Oct 25, 2026 sets out to fix.9
Where trace data may live decides most choices on record. Cue Labs and one developer's personal platform retired LangSmith for Langfuse in October 2026 to keep data in their own boundary, and Cresta and SuperDial self-host Langfuse for the same reason. monday.com and SonderMind stay on LangSmith for evals tied to release. Langfuse's MIT core self-hosts free; LangSmith self-hosts only on Enterprise.
What teams report
Left or turned down 1
NonBioS
NonBioS had rewritten its systems without LangChain, so it declined LangSmith as too tied to LangChain and chose Langfuse after a shortlist of Langfuse and Helicone (Mar 5, 2025). Session tracking that groups a run's calls, self-hosting and open source decided it.
“Langsmith appeared popular, but we had encountered challenges with Langchain from the same company, finding it overly complex for previous NonBioS tooling. We rewrote our systems to remove dependencies on Langchain and chose not to proceed with Langsmith as it seemed strongly coupled with Langchain.”
Amit, Author of the NonBioS engineering post, NonBioSNonBioS, , write-up
Adopted 4
monday.comOn LangSmith
monday.com wraps each Feed agent run in LangSmith's traceable() and runs an LLM judge in CI to check for regressions before production deployment (Sep 9, 2026).
“We use an LLM as a judge with specific metrics for each part, and run it on the CI to make sure we have no regressions before deploying to production.”
Kamie Shami-Schnitzer, Feed agent builder, monday.commonday.com Engineering, , write-up SonderMindOn LangSmith
SonderMind's clinical experts review production conversations of its mental health coach in LangSmith, and a script turns each missed disclosure they mark into an eval case (AI Engineer World's Fair 2026, June 29 to July 2; talk posted Jul 25, 2026).
“A clinical SME reviews a production conversation in LangSmith, marks the missed disclosure with an expected observation, and the script writes it to the eval set.”
Akele Reed and Dave Revere, Principal and Staff AI Engineers, SonderMindAI Engineer, , recording CrestaOn Langfuse
Cresta self-hosts Langfuse to trace its multi-service agent pipelines, one instance per Kubernetes cluster and one Langfuse organization per customer, with personal data redacted before recording (Mar 26, 2026).
“We self-host Langfuse rather than using a managed SaaS offering.”
Cresta Engineering, Engineering team, CrestaCresta, , write-up SuperDialOn Langfuse
SuperDial self-hosts Langfuse for all logging and observability of its voice agents because the calls carry healthcare data (AI Engineer Summit, February 2025; talk posted Apr 18, 2025).
“And then all of our logging and observability, we self-host Langfuse, and we self-host these things also because these are like healthcare calls.”
Nik Caryotakis, Staff Software Engineer, SuperDialAI Engineer, , recording
Choosing between LangSmith and Langfuse
Choose LangSmith if
- Release gates should run on your tracing vendor's experiments: monday.com runs an LLM judge in CI before each production deploy of its Feed agent (Sep 9, 2026), and SonderMind's clinicians turn annotated production conversations into eval cases.
- Your agents run on LangGraph, and you want Studio and LangSmith Deployment on the same platform as your traces.
- You want to choose retention per tracing project, with short-lived base traces priced below the ones you keep.
- You want LangChain to host your agents on Serverless or Dedicated deployments, or to run Agent Servers in your own VPC under a managed control plane.
Choose Langfuse if
- Traces must stay inside your own network without an Enterprise contract: Cresta self-hosts one Langfuse per Kubernetes cluster (Mar 26, 2026), and SuperDial self-hosts it because its calls carry healthcare data.
- You must self-host and want the full core platform free under the MIT license, on Docker Compose or Kubernetes, with your team running the collector, as GAIA learned when its own collector produced most of its error alerts.
- Many people need access and you want a flat platform fee with unlimited users from the US$29 Core plan up.
- Your code has left LangChain and you want a tracer with no tie to it, the reason NonBioS gave for ruling LangSmith out (Mar 5, 2025).
- You want cloud data access for up to 3 years on the Pro plan, or a data region in Japan.
Side by side
| Dimension | LangSmith | Langfuse |
|---|---|---|
| What teams reportNamed teams' own pages and talks. GILD found no team that compared both on the same traffic and published the numbers. | monday.com traces its Feed agent with traceable() and runs an LLM judge in CI before each deploy (Sep 9, 2026), and keeps one parent trace per Infra agent run (Oct 26, 2026). SonderMind's clinicians mark missed disclosures in LangSmith and each becomes an eval case (Jun 2026). GAIA ran LangSmith beside Langfuse after its development workspace hit the monthly trace limit (May 28, 2026).1238 | Cue Labs (Oct 9, 2026) and one developer's personal platform (Oct 11, 2026) retired LangSmith for Langfuse to keep data in their own boundary. Cresta self-hosts one instance per Kubernetes cluster (Mar 26, 2026); SuperDial self-hosts it for healthcare calls (Feb 2025). NonBioS chose it over LangSmith, which it found tied to LangChain (Mar 5, 2025).456710 |
| 200,000 traces a month, 10 users, kept 30 daysGILD's calculation from published rates, Oct 28, 2026. Inputs: 200,000 traces a month, 10 observations and 1 code score per trace, 10 users, at least 30 days retrievable, no judge calls, 1 LSU at US$1. | US$1,840 a month on Plus: US$390 for 10 seats, US$950 for 190,000 base traces past the 10,000 included at 0.005 LSU, and US$500 to move all 200,000 to the 180-day tier at 0.0025 LSU, since base traces last 14 days. At base retention alone, US$1,340.1113 | US$369 a month on Pro: US$199, plus US$72 for 900,000 units at US$8 per 100,000 and US$98 for the next 1.4 million at US$7, for 2.4 million units. Self-hosting the MIT core has no license fee; servers and operations are yours.2022 |
| PlansList prices in US dollars. LangSmith charges per seat; Langfuse plans include unlimited users from Core up. | Developer US$0 for 1 seat with 5,000 base traces a month; Plus US$39 per seat per month with 10,000 base traces a month; Enterprise custom11 | Hobby free with 50,000 units a month and 2 users; Core US$29 a month; Pro US$199 a month; Enterprise US$2,499 a month20 |
| Billing unitAn agent run with 10 observations and 1 score is 1 LangSmith trace and 12 Langfuse units, so compare on your own trace shape. | A trace: one execution of the application, including every run inside it, up to 25,000 runs13 | Units: each trace, each observation inside it, and each score count as one unit22 |
| Usage price | 0.005 LSU per base trace past the included volume, or US$5 per 1,000, plus 0.0025 LSU, or US$2.50 per 1,000, to upgrade a trace to extended retention; 1 LSU is US$111 | 100,000 units included on Core and above, then US$8 per 100,000, falling to US$6 per 100,000 above 50 million20 |
| Retention | Base traces 14 days, extended traces 180 days, chosen per tracing project; online evaluators and automation rules upgrade the traces they touch unless you opt out13 | Data access for 30 days on Hobby, 90 days on Core, and 3 years on Pro; project retention policies from Pro up, minimum 3 days2023 |
| Self-hosting | Enterprise add-on, on Kubernetes with ClickHouse, PostgreSQL, Redis, and blob storage16 | Free under the MIT license with all core features; Enterprise adds project RBAC, retention policies, and audit logs. GAIA's self-hosted collector caused 2,359 of the 3,890 Sentry events it logged over 30 days (Oct 25, 2026).92124 |
| Cloud regions | GCP us-central1, europe-west4, and australia-southeast1, and AWS us-east-21415 | US, EU, Japan, and a HIPAA region; US and HIPAA in AWS us-west-2, EU in eu-west-125 |
| Agent hosting | LangSmith Deployment: Serverless or Dedicated Agent Servers in LangChain's cloud, hybrid, self-hosted, or standalone, metered in LSUs18 | Not part of Langfuse's plans; it traces agents that run elsewhere20 |
| SSO and access control | Google and GitHub sign-in on Developer and Plus; custom SSO, custom roles, and ABAC on Enterprise1117 | Enterprise SSO and project RBAC through the US$300 a month Teams add-on on Pro, included on Enterprise20 |
| Compliance | SOC 2 Type 2, HIPAA, and GDPR; DPA on request15 | SOC 2 Type II and ISO 27001 reports, a HIPAA-ready region, and a GDPR DPA25 |
| Ownership | LangChain, which also maintains LangChain, LangGraph, and Deep Agents12 | ClickHouse, which acquired Langfuse in January 2026 and states no planned licensing changes26 |
What practitioners report, side by side
Each product's figures come from its own practitioners, for the group you pick.
Overall signal
All practitioners rate Langfuse +28, 11 points above LangSmith at +17. Most of the gap is in complaints, with 20 of LangSmith's 96 sources negative against 10 of Langfuse's 69.
The score counts the authors from all practitioners that take a side, one vote each with their latest view, and is how far positive votes lead negative ones, per 100 votes, so LangSmith's +17 rests on votes split 35 positive, 38 mixed and 19 negative. Each of Langfuse's authors wrote one source here, so each source is one vote.
A lead of a fifth of the votes reads positive and a lead of half strong positive, with the same lines below zero. The dots show each source, positive, mixed, negative. Run a pilot on one production app to measure which failures it surfaces and how long tracing them takes.
Footprint on the AI Application Platform stack
They differ on four of 9 layers. Coverage is each product's placement on this stack, as its product page shows it. Core is a job the product owns; Supports and Touches are jobs it helps with. Coverage says nothing about how well a product does them.
| Layer | LangSmith | Langfuse |
|---|---|---|
| Model | ||
| Model gateway | Supports | Touches |
| Build | ||
| Orchestration | Supports | Touches |
| Tool use | Supports | Touches |
| Operate | ||
| Cloud deployment | Supports | Touches |
The 5 layers where they match
| Layer | LangSmith | Langfuse |
|---|---|---|
| Model | ||
| Model APIs | Touches | Touches |
| Data | ||
| Retrieval | Touches | Touches |
| Evaluate | ||
| Evaluation | Core | Core |
| Operate | ||
| Observability | Core | Core |
| Guardrails | Touches | Touches |
What practitioners praise
Among all practitioners, Langfuse alone has three praise themes that reach 5 sources from 3 authors. Integrations & Setup and Open Source & Self-Hosting lead, at 11 of 11 and 8 of 8. Of the two themes both carry, Datasets & Evaluations differs most. It is praised in 13 of the 22 LangSmith sources that discuss it, against 2 of 7 for Langfuse. Quick Setup and Production Monitoring reach 5 sources from 3 authors only for LangSmith, at 11 of 11 and 4 of 14. Langfuse needs a pilot on the same production traffic before you choose on either.
| Praised for | LangSmith | Langfuse |
|---|---|---|
| Trace-Level Debugging | 28 of 35 | 13 of 14 |
| Datasets & Evaluations | 13 of 22 | 2 of 7 |
| Quick Setup | 11 of 11 | Not a theme in its sources |
| Production Monitoring | 4 of 14 | Not a theme in its sources |
| Integrations & Setup | Not a theme in its sources | 11 of 11 |
| Open Source & Self-Hosting | Not a theme in its sources | 8 of 8 |
| Prompt Management | Not a theme in its sources | 4 of 7 |
Each count is the sources that praise a theme, or raise it as a concern, out of those that discuss it. A row pairs the two products when both carry a theme under the same or an equivalent name.
A count ending in "so far" has under 5 sources from 3 authors among all practitioners. "Not a theme in its sources" means GILD found no theme of that name in that product's sources.
What practitioners raise as concerns
Among all practitioners, LangSmith alone has five concern themes that reach 5 sources from 3 authors. UI & SDK Reliability and Seat & Trace Pricing lead, at 23 of 25 and 12 of 14. Langfuse alone has four concern themes that reach 5 sources from 3 authors. Self-Hosting Operations and SDK & UI Reliability lead, at 14 of 15 and 12 of 12.
| Raised as a concern | LangSmith | Langfuse |
|---|---|---|
| UI & SDK Reliability | 23 of 25 | Not a theme in its sources |
| Seat & Trace Pricing | 12 of 14 | Not a theme in its sources |
| LangChain Coupling | 8 of 9 | Not a theme in its sources |
| Self-Hosting & Data Control | 8 of 12 | Not a theme in its sources |
| Closed Source & Vendor Dependence | 4 of 6 | Not a theme in its sources |
| Self-Hosting Operations | Not a theme in its sources | 14 of 15 |
| SDK & UI Reliability | Not a theme in its sources | 12 of 12 |
| Upgrades & Migrations | Not a theme in its sources | 12 of 13 |
| Feature Gaps | Not a theme in its sources | 6 of 6 |
26 sources
- monday.com Engineering: Building a Robust Harness for agent in production, Feed agent case study (Sep 9, 2026)
- monday.com Engineering: How We Gave AI Agents the Keys to Production and Still Sleep at Night (Oct 26, 2026)
- Akele Reed and Dave Revere (SonderMind): Evals Driven-Development, AI Engineer World's Fair 2026 (Jun 2026)
- Cresta: Observability for AI Agents, Tracing Multi-Service LLM Pipelines with Langfuse (Mar 26, 2026)
- Nik Caryotakis (SuperDial): Voice AI, Your Bot Isn't Special, AI Engineer Summit 2025 (Feb 2025)
- Cue Labs, cue-app pull request #513: consolidate LLM observability on Langfuse, retire LangSmith (Oct 9, 2026)
- retrieva-backend pull request #94: drop LangSmith, Langfuse as the single LLMOps backbone (Oct 11, 2026)
- GAIA pull request #720: self-hosted Langfuse alongside LangSmith (May 28, 2026)
- GAIA pull request #1510: keep trace-export failures visible (opened Oct 25, 2026)
- NonBioS: The NonBioS LLM Observability Pick (Mar 5, 2025)
- LangSmith plans and pricing
- LangSmith platform
- LangSmith docs: Usage and billing (retention tiers and rate limits)
- LangSmith docs: Cloud (SaaS) architecture and regions
- LangSmith docs: Regions FAQ
- LangSmith docs: Self-hosted LangSmith
- LangSmith docs: LangSmith for Enterprise
- LangSmith docs: LangSmith Deployment
- LangSmith docs: Studio
- Langfuse pricing
- Langfuse self-hosted pricing
- Langfuse docs: Billable units
- Langfuse docs: Data retention
- Langfuse docs: Self-host Langfuse
- Langfuse security and compliance
- Langfuse blog: Langfuse joins ClickHouse (Jan 16, 2026)