Prices verified 30 October 2026

Langfuse vs LangSmith: pricing, hosting and fit.

The two most-compared LLM tracing platforms, side by side. The short version: Langfuse is open source, self-hostable and much cheaper per run at volume; LangSmith is the tighter fit if you build on LangChain or LangGraph and spend most of your time in offline evaluation. Below is the detail, a worked bill at a million runs, and where neither one is the right tool.

Written by Agent Ping, which competes with both for production monitoring and says so in the last section. Prices read from each vendor's pricing page on 30 October 2026.

Side by side, on the things that decide it.

comparison Oct 2026
Dimension Langfuse LangSmith
Pricing model Plan fee plus graduated unit overage, no seats Per seat plus per-trace usage metered in LSUs
Free tier Hobby: 50,000 units a month, 30 days of data, 2 users Developer: 5,000 base traces a month, 1 seat
Entry paid plan Core, $29 a month with 100,000 units Plus, $39 per seat per month with 10,000 base traces
Billing unit A unit: each trace, observation or score A trace: one whole run tree, however many calls
Overage $8 per 100k units, falling to $6 above 50M $5 per 1,000 base traces, $7.50 extended
Data retention 90 days on Core, 3 years on Pro ($199) 14 days base, up to 180 days extended
Self-hosting Free, MIT licence, Docker Compose or Kubernetes Enterprise only (self-hosted or hybrid)
Framework fit Any: SDKs, OpenTelemetry, a long integration list LangChain and LangGraph first-class; others via SDK and OpenTelemetry
Evaluation and datasets Yes: datasets, LLM-as-judge, human annotation Yes: datasets, evaluators, online evaluation, mature
Prompt management Yes, versioned prompts with a playground Yes, prompt hub and playground
Ownership ClickHouse, since January 2026 LangChain Inc.
Production alerting Not a focus Not a focus

A worked bill at a million runs a month.

The unit is what decides the price. LangSmith's trace is generous (one run, however many calls) but costs $5 per thousand. Langfuse's unit is small (every observation counts) but costs $8 per hundred thousand, falling with volume. Two workloads, both at a million runs a month, on the cheapest paid plan of each.

monthly bill list prices
Workload Langfuse Core LangSmith Plus, 3 seats
Chatbot: 1M turns, one model call each (3 units, 1 trace) 3M units: $29 plan, 900k at $8 and 2M at $7 per 100k. About $241 1M traces: $117 seats, 990k at $5 per 1,000. About $5,067
Tool-using agent: 1M runs, six model calls and four tools each (11 units, 1 trace) 11M units: $29 plan, $72 band one, $630 band two, 1M at $6.50. About $796 1M traces: $117 seats, 990k at $5 per 1,000. About $5,067
Same agent, traces kept for 180 days Pro plan for 3 years of access: add $170. About $966 Extended traces at $7.50 per 1,000. About $7,542

LangSmith's bill does not change between the two workloads because the trace count does not. Langfuse's triples because the unit count does. Below about 50,000 runs a month the gap closes, and for a team of ten doing a few hundred evaluation runs a day the seat line is the whole LangSmith bill and the comparison flips. Full workings are in the LangSmith pricing and Langfuse pricing posts.

Which one, and when.

Pick Langfuse if

  • You want to own the data, or you have a residency requirement, and you have someone to run ClickHouse, Postgres, Redis and object storage.
  • Your run volume is high relative to your head count: the bill is units, not seats.
  • Your stack is not LangChain, or is several frameworks at once.
  • You are comfortable with the roadmap now sitting inside ClickHouse.

Pick LangSmith if

  • You build on LangChain or LangGraph and want instrumentation that is on by default.
  • Your main loop is offline evaluation on curated datasets, and the playground and prompt hub are daily tools.
  • Your team is larger than your trace volume: many people, modest production traffic.
  • You want deployment and agent hosting from the same vendor (LangSmith Deployment and Engine bill in the same LSUs).

Pick neither, or add a third, if

  • The question is what each customer cost you this month, not what a trace looked like. Both can tag runs; neither does per-customer attribution and alerting out of the box.
  • Your agents run on a schedule and you need to know when one silently stops. Tracing tools record what happened; absence of a trace is not an event they alert on.
  • You want every production run scored against a rubric without building the evaluation pipeline, and a flat bill with no unit to count.

That third job is what Agent Ping does, on a flat tier with named limits, and it runs beside either of these for teams that keep one for the development loop. The Agent Ping vs Langfuse and Agent Ping vs LangSmith pages set that out, and the Langfuse migration guide covers the run-both option.

Is Langfuse cheaper than LangSmith?
At production volume, by a wide margin. A million agent runs a month at eleven units each is about $800 on Langfuse Core; the same million runs is a million LangSmith traces at $5 per 1,000, about $5,000 before seats. At hobby volume both are free. The exception is a team of many people doing a small number of traces, where LangSmith's seats cost less than Langfuse's unit overage would.
Is Langfuse still open source after the ClickHouse acquisition?
Yes. ClickHouse acquired Langfuse on 16 January 2026 and stated that the core product remains open source under the MIT licence and that Langfuse Cloud continues as a standalone service. The Langfuse team moved to ClickHouse and says it will keep building the product.
Can I self-host LangSmith?
Only on the Enterprise plan, which offers self-hosted and hybrid deployments on a custom contract. Developer and Plus are cloud-only. Langfuse can be self-hosted for free on any plan with the published Docker Compose and Kubernetes templates.
Which is better for LangChain and LangGraph?
LangSmith. It is built by the same company and instruments LangChain and LangGraph with a single environment variable, with the trace tree, playground and evaluation flows shaped around those frameworks. Langfuse has a LangChain callback handler and LangGraph support that work well, but it is a general tool.
Do Langfuse and LangSmith do production monitoring?
Both are tracing and evaluation platforms first. Neither pages you when a scheduled agent stops running, attributes cost per customer out of the box, or scores every production run against a rubric without you building the pipeline. Teams that need those add a production tool such as Agent Ping beside whichever of the two they keep for the development loop.

More comparisons, and the pricing posts behind them.

LangSmith alternatives → Langfuse alternatives → LangSmith pricing explained → Langfuse pricing explained → Agent Ping vs Langfuse → Agent Ping vs LangSmith →