Langfuse vs LangSmith Pricing: Seats, Agent Traces and Retention Costs

An observability subscription buys debugging records, not cheaper model calls

Start with Langfuse Core for a cost-led team that needs collaborative tracing and can work within its selected feature and data-access arrangement. Evaluate LangSmith when its actual debugging and evaluation workflow saves enough work to justify the seat and trace expense.

The job is to understand a failed answer, a slow request or an unexpectedly expensive agent run. Logs should show the inputs, model calls, tool results and output that let a developer fix the problem. A dashboard that draws a trace but omits the context needed to reproduce the failure has not finished that job.

A production cost-tracking discussion asked which tools expose actual LLM spending. Keep the two invoices separate: a tracing subscription records model usage; it does not pay for the model calls it records.

One agent run is not one billable Langfuse unit

Langfuse's unit definition counts traces, observations and scores. For the workload used here, each application run creates one trace, five observations and one score:

1 trace + 5 observations + 1 score = 7 billable units
100,000 runs × 7 = 700,000 units
1,000,000 runs × 7 = 7,000,000 units

The observation and score counts are explicit workload inputs. A different instrumentation design changes them. An agent that adds tool calls or retries can create more observations without creating another customer-visible task.

LangSmith's trace meter follows its own execution definition. Use the product's tracing structure and count the records you actually ingest. Do not translate a Langfuse unit into a LangSmith trace one for one merely because both appear in a developer dashboard.

The team-size cost starts differently

The figures follow Langfuse's pricing and LangSmith's pricing, checked October 9, 2026. Do not multiply LangSmith's 10,000 included traces by the seat count. More seats add collaboration cost, not a documented separate 10,000-trace allowance for each person.

Put required identity and access controls on a separate procurement line. “Unlimited users” does not establish that every enterprise control is included in the $29 plan.

The modeled monthly bills at 100,000 and one million traces

This scenario uses seven Langfuse units per trace as defined above. LangSmith Plus uses its plan's 10,000 included base traces and the published $0.005-per-base-trace usage rate. Evaluator model calls, deployments and additional products are outside these totals.

Langfuse's graduated rate matters. For seven million units, the Core calculation is $29 plus $72 for the 900,000 units between the included 100,000 and one million, plus $420 for the next six million. Charging every additional unit at the first paid rate would overstate the bill.

For LangSmith's 100,000-trace example, 90,000 traces remain after the included allowance. At $0.005 each, usage adds $450. Add $117 or $390 of seats to obtain the table's totals. The products have different retention arrangements, explicitly shown in the column headings.

Longer retention is another meter

LangSmith distinguishes 14-day base traces from 180-day extended traces. Its current pricing adds $0.0025 per trace upgraded to extended retention. Upgrading all 100,000 traces adds $250; upgrading all one million adds $2,500.

Decide which records need the longer window. Retaining a small set of failed or reviewed runs is a different budget from extending every successful request. Langfuse Core's 90-day access and LangSmith's 180-day extended retention are also different requirements, not interchangeable versions of “long history.”

If your team needs longer Langfuse history access, price the relevant plan rather than treating Core as an unlimited archive. Access duration, deletion policy and exports deserve separate checks.

Evaluation can add records, model charges and retention changes

In Langfuse, a score is another billable record. An LLM judge can also create a model-provider charge. Adding evaluation to every trace changes more than the appearance of the dashboard.

LangSmith's administration guide describes the retention-extension settings attached to new evaluators and rules. Do not assume that every evaluation leaves the original retention charge unchanged. Conversely, ordinary UI feedback, notes or manual queue actions should not be described as universally triggering upgrades under today's rules.

Test one configured evaluation rule and inspect the resulting records and usage. Keep evaluator inference, tracing and retention on distinct budget lines. Turning a judge on for every request without measuring these effects is an easy way to make a small tracing purchase into a larger production bill.

Sampling reduces observation expense, not the cost of completed model calls

Langfuse's cost guidance includes reducing unnecessary recorded observations and sampling. Apply a deliberate policy: retain the records needed to diagnose failures and inspect quality while reducing redundant successful-run telemetry.

Do not log less and claim the underlying AI application became cheaper. If a model call still ran, its provider can still bill it. Observability can reveal an expensive loop and help you remove it; skipping the record does not remove the loop itself.

Preserve enough context to connect a record to the code and prompt configuration that produced it. Otherwise a cheaper log stream can become less useful precisely when a production failure requires investigation.

Self-hosting is a different operating commitment

Langfuse's self-hosted option changes control over the deployment and data, but someone still owns capacity, upgrades, backups and restoration. Thoughtworks's Technology Radar describes experience with Langfuse and the complexity of its newer self-hosted architecture. A free software license is not a zero-cost production service.

Estimate the infrastructure and the staff responsibility before comparing self-hosting with either managed subscription. Use the selected deployment's security and data controls to answer where records live and how sensitive fields are handled.

For the modeled team, Langfuse has the clear subscription-cost advantage

Langfuse Core is substantially cheaper in the stated seven-unit workload, even after counting its observations and scores correctly. It is the first cost-led evaluation for a team that can use its selected controls and history-access window.

Choose LangSmith when its specific tracing and evaluation workflow earns the additional expense in your own debugging work. Check the record that lets a developer explain a failed run, the time to reproduce it and the cost of the retained evidence. That is a stronger purchase test than comparing two feature lists or treating unlike billing units as equal.

Is a Langfuse unit the same as a trace?

No. Traces, observations and scores each contribute to billable units.

Does buying ten LangSmith Plus seats include 100,000 traces?

Do not multiply the plan's stated 10,000 included traces by seat count.

Does less tracing reduce the model-provider bill?

Not by itself. It reduces recorded observability work; completed model calls remain separate.