Langfuse vs LangSmith for Reviewing LLM Traces

When an LLM feature gives a bad answer, the final response rarely explains the failure. You need to see which prompt was sent, what retrieval returned, where a tool failed and whether a retry changed the result.

Langfuse and LangSmith both address that inspection problem. The useful choice for a small engineering team depends on its instrumentation, review process, history requirements and willingness to operate infrastructure.

Quick verdict: Start with Langfuse when an open-source self-hosting route or Cloud collaboration without per-seat pricing is central to the decision. Put LangSmith on the shortlist when its supported integrations and trace-to-feedback workflow fit the team's existing development process. Neither choice establishes better model quality by itself.

Basis: An editorial comparison of official documentation and public prices, checked October 21, 2026. We did not install either platform, send production traces, benchmark overhead or test detection accuracy. The cover is a conceptual illustration.

Compare the same debugging task

Consider a proposed evaluation: a support assistant retrieves documents, calls a model, invokes a tool and sometimes retries. An engineer must determine whether a wrong answer came from retrieval, the tool result or the final generation. A second reviewer should be able to find the same evidence later.

Use the same application version and permitted, sanitized requests for both candidates. Keep these criteria fixed:

• Can the reviewer follow parent and child operations through a request?
• Are errors, timing, model identity and relevant inputs visible?
• Can feedback be attached to a reproducible example?
• Can the required people inspect the data for the required period?
• What does the actual instrumentation generate for billing?

This is a decision framework, not a report of a completed head-to-head test.

Langfuse vs LangSmith at a glance

The feature rows describe documented capabilities, not equal coverage for every SDK or plan. Confirm the exact integration and permissions your reviewers will use.

Inspection and integration: choose by the captured evidence

Langfuse's observability documentation describes a trace with nested observations, including prompts, responses, timing and intermediate work. It also connects trace inspection with scores and evaluation. This suits a team that wants the same request context to support debugging and later quality checks.

LangSmith's observability overview covers trace inspection, monitoring and feedback. Its data model distinguishes a run for an individual operation from a trace grouping the runs for that operation; threads can group multiple turns. Manual instrumentation is available alongside framework integrations.

Do not choose LangSmith solely because the application uses LangChain, or exclude it solely because it does not. Likewise, an OpenTelemetry integration is a reason to investigate Langfuse, not proof that every field will arrive with the meaning you expect.

For either candidate, check one successful request and one failed request. Confirm that the tool call is attached to the right parent, retries remain distinguishable and the error is visible at the relevant step. A visually complete tree can still contain missing payloads or incorrect identifiers. Record what the integration actually captures before adding more dashboards.

If the application uses a model gateway, keep gateway behavior in scope. Our OpenRouter prototype review explains why model identity, provider routing and spending controls need separate checks. An observability platform cannot reconstruct a provider decision that your integration never records.

Hosting: software availability and operating responsibility

Langfuse's self-hosted pricing lists the core platform under the MIT license, including observability, evaluation, prompt management and datasets. Enterprise capabilities remain a separate offering. That makes an OSS deployment a concrete option for a team that needs infrastructure control and has someone to maintain it.

Free software is not a free operated service. Budget application and database capacity, storage, backups, upgrades and incident handling. Check feature licensing individually rather than assuming that an OSS installation includes every Cloud administration control.

LangSmith's platform setup guide offers Cloud, BYOC and self-hosted options; BYOC and self-hosted are Enterprise. For a small team without an infrastructure owner, managed Cloud is the more relevant starting comparison. A requirement to keep observability data in your infrastructure changes the commercial discussion.

Choose by constraint: If self-hosting is mandatory and an Enterprise contract is out of scope, Langfuse OSS is the practical candidate of these two. If managed operation is acceptable, compare the captured traces and review workflow before treating hosting as the deciding factor.

Prices and limits: units are not interchangeable

These are published USD monthly rates, checked October 21. They are not annual-payment equivalents or all-inclusive project quotes.

Sources: Langfuse Cloud pricing and LangSmith pricing. Paid usage beyond allowances and optional services can add cost. Langfuse Core currently lists additional units at $8 per 100,000 in its first paid usage tier, with lower rates at higher volumes. Confirm the applicable calculator and checkout terms.

The key distinction is the meter. Langfuse's billable unit definition adds traces, observations and scores. A single application request can therefore produce several units. LangSmith's included volume is expressed in base traces; do not compare that headline number directly with Langfuse units or assume that a child run is another included root trace.

Here is a hypothetical monthly workload, not measured usage: 20,000 requests, each producing one Langfuse trace, five observations and one score. That is 140,000 units before any additional platform-generated work. Under the listed Core first-tier rate, the illustration is $29 plus $3.20 usage, or $32.20. It excludes model inference, add-ons, taxes and any other billable activity.

For three LangSmith Plus seats, the seat component alone is $117/month. Add the applicable trace usage and any optional services; $117 is not the full bill. This does not establish that one product is cheaper for your workload: the captured volume, retained trace tiers and enabled features must be measured first.

LangSmith's billing guide says a personal Developer organization is capped at 5,000 monthly traces until a card is added; adding billing enables excess usage charges. Review workspace usage limits before increasing ingestion. A free plan is not a time-limited trial, and a paid plan is not unlimited telemetry.

History access is not the same as deletion

Langfuse's retention documentation explicitly distinguishes access windows from deletion. The Cloud window limits visible history; it does not itself delete the data. Configurable retention is available on Cloud Pro and Enterprise and on self-hosted Enterprise. Without a policy, self-hosted event data is kept indefinitely by default.

LangSmith's usage and billing guide distinguishes 14-day base retention from 180-day extended retention. It notes a September 14, 2026 change applying to new traces, with existing traces unaffected. The general concepts page also describes a 180-day SaaS period without that tier distinction. Use the detailed billing guide and verify your project's actual tier rather than assuming every trace has 180 days. Trace content is deleted after expiry, while some metadata may remain for billing and analytics.

For either product, define the real requirement: two weeks of debugging, a quarterly review or a documented deletion deadline. A saved dataset example and its source trace are different records; do not assume that saving one preserves the other's full execution history.

Sensitive payloads require an explicit capture policy

Both products document controls before data reaches the tracing backend. Langfuse supports masking in the export path. LangSmith supports hiding inputs, outputs and metadata or applying masking.

Decide which fields may be captured, then verify the exact SDK path, including errors and nested tool calls. Redacting a prompt does not automatically redact an exception string or metadata value. Also inspect other exporters: configuring one backend's masking does not prove that a second telemetry destination receives the same filtered data. Self-hosting does not remove the need for those checks.

Our How to Review a First LLM Trace with Langfuse shows how to inspect one synthetic request, its nested operations and export completion.

Which should a small team choose?

Choose Langfuse first when OSS self-hosting is a hard requirement, or when two-person free access and paid collaboration without seat fees match the team's needs. Its trade-offs are unit growth from detailed instrumentation and the need to distinguish history access from retention controls.

Choose LangSmith first when the supported integration and run-review workflow match the application, and the team accepts seat-based collaboration or already has an appropriate Enterprise agreement. Its trade-offs include paid seats for team use and the need to budget trace tiers and optional services separately.

Delay the purchase if neither candidate captures the step that causes the failure. Add or correct instrumentation first. For a tiny prototype with occasional failures, sanitized structured logs may be enough until the review burden justifies a dedicated platform.

Open the Langfuse and LangSmith tool pages, then evaluate the same sanitized requests against the five criteria above. Keep a short record of missing fields, review effort, generated billing volume and the required history window. Choose the service that satisfies those constraints with an acceptable operating cost.