Langfuse vs LangSmith: Self-Hosting and LangGraph Depth

Langfuse is the stronger fit when a team needs free self-hosted software or traces agents built across several frameworks. LangSmith is the stronger fit when LangGraph state inspection and replay are central to debugging. DataCamp’s independent comparison describes that divide while noting that both products support tracing and evaluation.

The cost decision is less obvious than the feature choice. Langfuse Cloud counts the observations and scores inside a trace; LangSmith charges for seats and traces, with an additional charge when traces receive extended retention. A team should compare a representative agent run and the history it needs to keep, rather than equating one request with one billable item on each platform.

Free self-hosting still means operating the stack

Langfuse self-hosting depends on application containers, databases, cache, and blob storage that the team must operate.

Langfuse’s MIT-licensed edition has no software license or usage fee for self-hosting. LangSmith offers self-hosted and hybrid deployment through Enterprise, for which it does not publish a standard price. If the requirement is to run the observability platform on infrastructure the team controls without an Enterprise contract, Langfuse provides the clearer route.

That route has a substantial infrastructure footprint. Langfuse’s sizing guide lists web and worker containers, PostgreSQL, Redis or Valkey, ClickHouse, and blob storage; its minimum for ClickHouse alone is two CPUs and 8 GiB of memory. Production cost also includes persistent storage, backups, upgrades, monitoring, and the time to respond when ingestion or queries slow down. The free license removes a vendor charge, not those operating expenses.

Ownership can be the point: the team chooses where the services run and operates the trace store alongside its other systems. For a small team without database operations capacity, Langfuse Cloud may be the more economical way to use the product, even though the self-hosted code is free. A self-hosted estimate should therefore use actual infrastructure and staffing costs, rather than treating the license price as the total.

LangGraph integration changes what debugging reveals

LangSmith’s advantage is the context around a LangGraph execution. Its native workflow exposes graph nodes and state transitions in Studio and supports inspection and replay. Langfuse can also trace a LangGraph run and display its agent graph, but its trace structure is designed to work across frameworks. LangSmith also accepts traces from applications outside LangGraph, so the difference is the depth of this particular integration, not whether other agents can be observed.

Consider an agent that chooses a tool, receives a result, and takes a different branch than expected. A nested trace can show the tool call, model response, timing, and error. State inspection adds the graph context needed to see what was available at a node and follow how execution progressed. That extra view has the greatest value when the team routinely investigates LangGraph branches or replays runs; a team operating several agent frameworks may value a common trace and evaluation layer more.

A matched run exposes evaluation and setup differences

A reproducible comparison can use one conditional LangGraph workflow with a tool call, a model response, and a quality score attached to the outcome. This is a proposed trial, not a claim that the platforms were tested here. Run the same application input and branch through each integration, then compare the recorded tool arguments, model output, graph context, and score.

For LangSmith, configure its API key and enable tracing, then run the graph and inspect the resulting trace and state in Studio.
For Langfuse, configure the project keys and endpoint, add its LangChain callback handler to the graph invocation, then inspect the resulting trace, nested observations, and score.
Put the failed outcome into each product’s dataset and evaluation workflow. Record which context is available without extra instrumentation and whether evaluation changes the trace’s retention or billable usage.

Both products support datasets and evaluations using model-based or code-based checks. The useful distinction is the path from a failed agent run to an explanation: LangSmith keeps that work close to LangGraph state, while Langfuse keeps scores and experiments alongside traces from other runtimes. An evaluator also matters to cost. A score adds a Langfuse unit; in LangSmith, an online evaluator configured to extend retention can move the scored trace into the longer-lived tier.

Historical access and trace retention are different purchases

LangSmith’s plans list Plus at $39 per seat with 10,000 included monthly base traces, reserve self-hosted and hybrid options for Enterprise, and give base and extended traces retention periods of 14 and 180 days, respectively. The Developer plan offers one free seat and 5,000 included base traces. A LangSmith trace represents an application execution and can contain many individual model and tool steps.

Langfuse’s cloud pricing lists a free Hobby tier with 50,000 monthly units and 30 days of historical access, Core at $29 a month with 100,000 units and 90 days, and Pro at $199 a month with 100,000 units and three years. The paid tiers allow unlimited users and charge for units beyond the included allowance. A Langfuse unit is a trace, an observation, or a score, so deeper instrumentation increases usage even when the number of agent executions stays fixed.

Langfuse calls its plan windows historical access; those windows should not be read as promises that data is deleted at their end. LangSmith describes retention for individual traces, and an extended trace carries an upgrade charge. The distinction matters if a team needs older runs for regression analysis or has a separate data deletion requirement. In LangSmith, retention extension is enabled by default for new online evaluators and automation rules, although it can be disabled for those actions.

What a small team and a busy agent would pay

For a hypothetical five-person team, 20,000 runs produce 100,000 Langfuse units and 20,000 LangSmith traces, yielding different modeled monthly charges.

BenchClaw’s reproducible pricing extraction found LangSmith’s published tooltip rate of 0.005 LSU per additional base trace and 0.0025 LSU per extended-retention upgrade; at $1 per LSU, that is $5 and $2.50 per 1,000 traces, respectively. Its extraction reads published rates rather than an observed customer invoice. Langfuse’s paid cloud overage begins at $8 per 100,000 units and falls at higher published volume tiers.

Take a hypothetical five-person team with 20,000 agent runs in a month. If each run produces one trace, three observations, and one score, Langfuse records 100,000 units, leaving Core at its $29 base price. The same workload gives LangSmith 20,000 traces: five Plus seats cost $195, and the 10,000 traces above the included allowance add $50, for $245 before retention upgrades or other metered services. The example assumes every LangSmith trace stays in the base tier.

For a hypothetical high-volume agent, assume 500,000 monthly runs, each producing one trace, six observations, and one score. That is four million Langfuse units. Applying its graduated rates yields about $311 on Core or $481 on Pro. With the same five seats, LangSmith Plus would be about $2,645 for 500,000 base traces. These are arithmetic estimates from the stated run shapes and list prices, excluding model usage and any self-hosted infrastructure. Core’s shorter historical access, Pro’s longer access, and LangSmith’s base retention mean the three totals buy different lengths of usable trace history.

Trace depth is the variable most likely to change the Langfuse estimate; the number of collaborators and the share of traces upgraded for retention can change the LangSmith estimate. If graph-state replay is a daily debugging need, its value may justify the LangSmith bill. If infrastructure control or a shared evaluation layer across runtimes is the lasting requirement, Langfuse’s deployment choices may matter more than the lowest modeled cloud price.