Langfuse Pricing
Langfuse pricing tells you exactly what tracing your agent costs. It never tells your customer what the agent was worth.
Langfuse's real numbers are simple once you cut past the third party writeups guessing at them: Hobby is free for fifty thousand units a month, Core is $29, Pro is $199, and Enterprise starts at $2,499, each verified straight from Langfuse's own pricing page. A 'unit' means a trace, an observation, or a score, and overage past the included allowance runs on a graduated per hundred thousand unit rate that drops as volume climbs. Self-hosting the open source version is genuinely free; a specific dollar figure for self-hosted Enterprise that shows up on a couple of other pricing pages doesn't appear anywhere on Langfuse's own site, which only lists it as custom, sales-quoted pricing. This page covers every tier in full, corrects that unverified number, and covers the one question none of Langfuse's plans, free or Enterprise, were built to answer: what did the agent actually do for the customer paying for it.
Verified against Langfuse's own pricing page
What Langfuse Cloud actually costs
Langfuse's cloud pricing runs four tiers, and every number below comes from Langfuse's own current pricing page rather than a list price someone forgot to update. Hobby costs $0 a month, covers fifty thousand units, holds data for thirty days, and caps a workspace at two users, with support handled through Langfuse's own GitHub community rather than a dedicated channel. Core costs $29 a month, keeps the same fifty thousand unit starting point but bills overage past a hundred thousand included units, extends retention to ninety days, and removes the two user cap entirely, unlimited seats included in the base price. Pro costs $199 a month, holds the same hundred thousand included units as Core but stretches retention out to three full years, the single biggest jump in the whole pricing table, and adds prioritized in-app support over Core's community-only tier.
Enterprise starts at $2,499 a month, and the jump from Pro isn't about the unit allowance, which stays at a hundred thousand included units same as Core and Pro. It's about what gets added around that allowance: audit logs, a SCIM API for automated user provisioning, a dedicated support engineer, and an uptime SLA available on request, plus custom volume pricing once a company commits to a yearly contract instead of paying month to month. A company that needs SSO or role based access control but isn't ready for the full Enterprise contract can add Teams to the Pro plan for $300 a month instead, a detail easy to miss since Langfuse's own pricing table lists it as a single add-on line rather than its own column next to the four main tiers.
Four tiers, one meter
The four Langfuse Cloud tiers, in plain numbers
Straight from Langfuse's own pricing page, not a third party estimate.
Hobby: $0 a month
Fifty thousand units, thirty day retention, capped at two users. Support runs through GitHub community threads rather than a dedicated channel.
Core: $29 a month
A hundred thousand included units before overage, ninety day retention, and unlimited users from this tier up. The step most teams past testing actually take.
Pro: $199 a month
Same hundred thousand included units as Core, retention jumps to three years. SSO and RBAC cost an extra $300 a month Teams add-on here, not included by default.
Enterprise: $2,499 a month
Audit logs, SCIM provisioning, a dedicated support engineer, and an SLA, all included at this price rather than billed as a separate add-on.
The number every tier bills against
How a 'unit' actually gets counted
Every tier above bills against the same meter, and Langfuse's own documentation defines it precisely: a unit is the sum of a trace, an observation, and a score, counted separately and added together. A trace is one complete agent request from the moment it starts to the moment it finishes. An observation is one individual step inside that trace, a single tool call, a single retrieval, a single model response. A score is one evaluation result attached to either. Langfuse's own worked example puts this in real numbers: a deployment generating twenty thousand and seventy traces, a hundred and nineteen thousand five hundred observations, and five hundred and sixty one scores in a month totals a hundred and forty thousand one hundred and thirty one billable units, well past what the trace count alone would suggest. Units that Langfuse's own features generate automatically, LLM as a judge scoring, annotation queues, and background experiments, count toward that same total, which is the detail that most often explains a bill that looks higher than the raw request volume a team expected.
Past the included allowance on any paid tier, Langfuse doesn't cut a workspace off; it keeps ingesting and bills the overage on a graduated scale. The next block, a hundred thousand up to a million units, costs $8 per hundred thousand units. From a million up to ten million it drops to $7 per hundred thousand, from ten million up to fifty million it drops again to $6.50, and past fifty million units a month the rate settles at $6 per hundred thousand, the same graduated schedule applying identically whether a workspace sits on Core, Pro, or Enterprise. One correction worth making here: a specific self-hosted Enterprise license figure, around $500 a month, shows up on a couple of third party pricing breakdowns of Langfuse, but that number doesn't appear anywhere on Langfuse's own self-hosting page, which lists Self-Hosted Enterprise only as custom pricing bundled with a separate ClickHouse commercial plan, reachable by contacting sales rather than checking a published rate card. Treat that $500 figure as unverified until Langfuse's own team confirms a real number for your actual usage.
Free to run, not free to fully operate
Self-hosting Langfuse: what's actually free and what isn't
Also verified straight from Langfuse's own self-hosting page, including the one figure other writeups get wrong.
Self-Hosted Open Source
Genuinely free under an MIT license. Unlimited units, no usage cap, every core feature, and even organization level RBAC and enterprise SSO included, features Cloud plans charge extra for. Support runs through the open source community only.
Self-Hosted Enterprise
Adds project level RBAC, audit logs, data retention policies, SOC 2 and ISO 27001 reports, and a dedicated support engineer. Priced custom, bundled with a separate ClickHouse commercial plan, quoted only through sales, no flat number published.
What the unit meter doesn't show
Langfuse's meter tells you what tracing cost. It tells your customer nothing
Every number on Langfuse's pricing page, the unit count climbing toward a plan's allowance, the graduated overage rate, the three year retention window on Pro, describes the same thing: what tracing your own agent costs you and your engineering team. That holds true at every tier, including the self-hosted Enterprise contract, since even audit logs and SCIM provisioning are still built for your own account administrators, not for an outside customer. Unlimited users on Core and up means unlimited seats for your own company, and even Langfuse's own multi-tenancy inside a workspace, organization level RBAC, unlimited projects, still assumes every project belongs to your team, not to a paying customer sitting outside it entirely.
If you sell an AI agent inside a product other companies pay for, that gap shows up the moment a customer asks what the agent actually did for their account this month, a question no unit meter or retention window answers. AiAgRe's Node SDK reads the same kind of agent events Langfuse already traces, through the same LangChain, LlamaIndex, and CrewAI integrations, and tags each one with an organization identity and a customer identity at the point of ingestion. That turns into white-labeled dashboard components your customer sees inside your own product, showing deflection rate and cost per resolution scoped so tightly that one customer's numbers never reach another customer's view. It doesn't replace Langfuse's trace level tracing, and most teams run both: Langfuse for what the agent did, AiAgRe for what that activity was worth to the person paying for it.
When a unit meter stops being the real question
Three questions Langfuse's pricing page was never built to answer
None of these show up in a unit count or a retention window. All three show up once real customers start asking.
Your customer asks what their account got, not how many units it used
A customer renewing a contract wants their own deflection rate and cost per resolution, scoped to their own traffic, not a walkthrough of a unit meter scoped to an account they can't log into. A clean trace reassures your engineers; it says nothing to the person deciding whether to renew.
Unlimited seats, zero customer tenants
Core, Pro, and Enterprise all remove the seat cap for your own team as you move up. None of the three add a scoped view for a customer outside your company, because Langfuse's multi-tenancy was built for your own projects, not for a second tenant layer underneath them.
The overage rate tracks volume. Nothing tracks value
The graduated $8 to $6 per hundred thousand unit rate answers what an agent's traffic costs as it grows. It says nothing about what that traffic was worth, the same gap a fuller look at AI agent monitoring tools as a category runs into past Langfuse alone.
FAQs
Langfuse pricing: frequently asked questions
Common questions from teams pricing Langfuse against their own usage before they commit to a tier, or before they add a second layer on top.
How much does Langfuse actually cost?
The Hobby plan is free for fifty thousand units a month, with thirty day retention and two users, support handled through GitHub's community forum. Core runs $29 a month for a hundred thousand units, ninety day retention, and unlimited users. Pro runs $199 a month for the same hundred thousand included units but three year retention, plus a $300 a month Teams add-on if a company needs SSO, role based access control, or a dedicated Slack channel. Enterprise starts at $2,499 a month, also for a hundred thousand included units and three year retention, and adds audit logs, a SCIM API, a dedicated support engineer, and an uptime SLA. Every figure above comes straight from Langfuse's own pricing page as of this page's publish date, worth reconfirming there before budgeting against it since usage based rates change.
What actually counts as one 'unit' in Langfuse's pricing?
Langfuse's own documentation defines a unit as the sum of three separate event types: a trace, an observation, and a score. A trace represents one full agent request from start to finish, an observation represents one individual step inside that trace such as a tool call or a model response, and a score represents one evaluation result attached to either. Langfuse's own worked example puts twenty thousand and seventy traces, a hundred and nineteen thousand five hundred observations, and five hundred and sixty one scores at a hundred and forty thousand one hundred and thirty one billable units for that month. Units generated automatically by Langfuse's own features, LLM as a judge scoring, annotation queues, and experiments, count toward the total the same as anything sent directly from an application, which is easy to miss until a bill arrives higher than the raw request volume would suggest.
What happens once a team goes over the included units?
Langfuse keeps ingesting data past the included allowance rather than cutting anything off, and bills the overage on a graduated scale that applies the same way across Core, Pro, and Enterprise. The next block, from a hundred thousand up to a million units, costs $8 per hundred thousand units. From a million up to ten million it drops to $7 per hundred thousand, from ten million up to fifty million it drops again to $6.50, and past fifty million units a month it settles at $6 per hundred thousand. A team running real production traffic through an agent should treat the $29 or $199 base price as a starting point rather than the real monthly number, since the overage tiers, not the base plan, usually decide what the invoice actually looks like once volume grows.
Is self-hosting Langfuse actually free?
Partially, and the distinction matters more than most writeups about it let on. The Open Source self-hosted tier is genuinely free under an MIT license, covers unlimited units with no usage cap, and includes every core feature plus organization level role based access control and enterprise SSO, features Langfuse charges extra for on its own Cloud plans. What it excludes is server side data masking, project level role based access control, configurable data retention policies, audit logs, and dedicated support, all of which move to Self-Hosted Enterprise. Langfuse's own self-hosting page lists that Enterprise tier as custom priced, additive to a separate commercial ClickHouse plan, and reachable only by contacting sales, with no flat monthly number published anywhere on the page. A specific dollar figure for that tier that circulates on a couple of third party pricing writeups doesn't appear on Langfuse's own site at all, so treat any exact self-hosted Enterprise number you see elsewhere as unverified until Langfuse's own sales team confirms it for your actual usage.
Is Langfuse cheaper than LangSmith?
At the free tier, yes, by a wide margin: Langfuse's Hobby plan covers fifty thousand units a month against LangSmith's Developer plan, which covers five thousand base traces before its own pay as you go pricing kicks in. Past the free tier the two get harder to compare directly, since a Langfuse unit and a LangSmith trace aren't the same thing and LangSmith's overage runs through its own LangChain Compute Unit at $1.50 and Storage Unit at $1 rather than Langfuse's per hundred thousand unit rate. The fuller side by side, including where Confident AI, Arize AI, and Braintrust fit into the same comparison, is on the LangSmith alternatives page.
Does any Langfuse plan show pricing or ROI data to my own customers?
No, on every tier including Enterprise. Every dashboard Langfuse renders, whether it's the unit meter counting toward a plan's included allowance or the trace level detail underneath it, lives inside a workspace scoped to your own account and built for your own engineers to read. Unlimited users on the Core plan and up means unlimited seats for your own team, not a scoped view for an outside customer. If you sell an AI agent inside a product other companies pay for, your own customer wants their own deflection rate and cost per resolution, isolated to their own traffic only, which sits outside what any tracing tool, Langfuse included, was built to show.
Can I run AiAgRe alongside Langfuse?
Yes, and that's the normal setup rather than a workaround. AiAgRe's Node SDK reads the same kind of underlying agent events Langfuse already traces, through the same LangChain, LlamaIndex, or CrewAI integrations, and tags each one with an organization identity and a customer identity at the point of ingestion. That turns into white-labeled dashboard components your own customer sees inside your product, scoped so tightly that one customer's numbers never reach another customer's view. Installing it doesn't touch how Langfuse ingests a unit or bills for it; the two SDKs read the same agent, one scoped to your engineers and billed per unit, the other scoped to your paying customers.
Does the $300 a month Teams add-on apply to every paid plan?
No, it's specific to the Pro plan. Core ships without SSO or role based access control at any price, and Enterprise already includes SCIM, audit logs, and dedicated support in its $2,499 base price without a separate add-on. The $300 a month Teams add-on exists specifically to bridge Pro up to Enterprise level access controls, SSO, RBAC, and a dedicated Slack channel, without moving a company onto the full Enterprise contract, and it's easy to miss on a first read of Langfuse's pricing table since it appears in a single line rather than its own column.
What would a mid-size workload actually cost on Core?
Take a workspace generating five hundred thousand units in a month, a reasonable number for an agent handling a few thousand real conversations once tool calls and observations are counted alongside traces. Core's base price of $29 covers the first hundred thousand units. The remaining four hundred thousand fall inside the first overage bracket, a hundred thousand up to a million units at $8 per hundred thousand, so four blocks at $8 add $32. The real bill for that month lands at $61 total, not the $29 sticker price on Langfuse's own pricing page, and that gap only widens as unit volume grows past a million and the graduated rate starts working in the workspace's favor instead.
Is the free Hobby tier usable for a live production agent?
For a small pilot, yes; for a real production workload, usually not for long. The fifty thousand unit allowance and thirty day retention window are generous for testing an integration or running a low-traffic beta, and the price is genuinely $0. The two user cap and GitHub-only community support are the two limits that tend to force an upgrade first, well before the unit allowance does, since most teams past a single-developer pilot need at least a handful of people looking at the same traces. Treat Hobby as a real starting point rather than a permanent home for a live agent serving paying customers, and budget for Core once a second engineer needs access.
Already paying to trace your agent? Now show your customer what it did.
Request access and we'll walk through how AiAgRe's embed tokens map onto the trace data Langfuse already produces.