Table stakesBuilding blockAI-native

Langfuse

Langfuse GmbH, Berlin (owned by ClickHouse, Inc. since 2026-01-16) ·langfuse.com· As of: 2026-07-30

Once you build language models into a product, you cannot see what the model was asked, what it answered, what the call cost, or whether a prompt change improved or degraded quality. Langfuse records every model call, versions the prompts and scores answers against fixed test cases.

Our verdict

If you do not record model calls, you cannot trace faults, cannot evidence quality and cannot attribute cost. That is a regression you prevent, not an advantage you build. The platform itself has become interchangeable, several vendors ship the same feature set. What gains value over time is your own body of test cases and evaluation runs, and that belongs to the operator, not to the tool. It can be exported and carried elsewhere. Hence mandatory kit, not an edge.

Who it fits

Solo

suitable

The free tier covers 50,000 units a month and two seats, with retention ending after 30 days. That is enough for a single application in trial, and the MIT-licensed core can also be self-hosted at no charge.

Mid-market

suitable

Twenty-nine dollars a month with unlimited seats and 90 days of retention suits a team running one or two AI features in production. The unknown is consumption, because every recorded event counts and multi-step chains produce many events per request.

Enterprise

suitable with caveats

Conditional, because the very parts a large organisation must have sit at the expensive end. Single sign-on and fine-grained roles cost an extra 300 dollars a month, audit logs and SCIM provisioning only arrive in the 2,499 dollar plan. When self-hosting, those same features sit behind a paid licence key whose price is not published.

Buy, switch or build

Off-the-shelf

Log model calls into the observability stack you already run

If you already run an observability platform, model calls can be filed there as ordinary traces. That saves a separate procurement, access rights are settled and the storage location is known. What is missing is any semantics for input, output, prompt version and score. You can see that a call was slow or expensive, but not whether the answer was usable. With AI features that is precisely the question that matters.

AI-native

One of the competing AI-native platforms

The category as a whole is AI-native and Langfuse is part of it. The difference between vendors barely lies in features any more, but in the licence, the storage location and who owns the vendor. Langfuse distinguishes itself through the MIT-licensed core and the selectable EU region. Choosing a closed platform instead gives up exactly that escape route and leaves you dependent on the vendor's export interface.

Build it yourself

Trace collection into your own database

The ingestion path is manageable: standard OpenTelemetry traces into your own database plus a plain view on top. That holds as long as only developers look at it. It breaks on the parts that carry the real effort, namely evaluation runs against fixed test cases, prompt versions with approval, and an interface where domain experts without coding skills can judge answers. Building that layer yourself costs more, permanently, than it returns.

Our advice

  • Solo: Free tier or self-hosting, both hold up.
  • Mid-market: Buy, pick the EU region, recheck consumption after four weeks.
  • Enterprise: Buy or self-host, but settle first what audit logging costs and who the contracting party is.

Who is behind it

Staying power: Funded

The legal entity is German, control has not been since January 2026. Operations today rest on a heavily funded acquirer rather than on a vendor that has been self-sustaining for years. The MIT-licensed core lowers the risk noticeably, because continued operation stays possible without the vendor.

  • Change of ownership: acquired by ClickHouse, Inc., announced on 16.01.2026; purchase price not disclosed · Source · As of 2026-07-30
  • Funding of the owner: USD 400 million Series D led by Dragoneer, announced the same day · Source · As of 2026-07-30
  • Legal entity and register: Langfuse GmbH, Oranienburgerstraße 91, 10178 Berlin, commercial register Amtsgericht Charlottenburg HRB 248821B, managing directors Aaron Katz and Clemens Rawert · Source · As of 2026-07-30
  • Adoption: more than 32,000 GitHub stars, more than 6 million Docker pulls, more than 50 million SDK installs per month, in use at 21 of the Fortune 50 and 129 of the Fortune 500 · Source · As of 2026-07-30
  • Certification: SOC 2 Type II and ISO 27001, audited annually, annual independent penetration tests · Source · As of 2026-07-30
  • Licence of the core: MIT, expressly confirmed by the acquirer to remain so after the takeover, self-hosting at production scale permitted · Source · As of 2026-07-30

Cost of leaving: Low

Three routes lead out and all of them are open. The core is MIT-licensed and can be self-hosted. Traces, observations and scores can be exported on a schedule to S3, Google Cloud Storage or Azure Blob Storage in Parquet, CSV, JSON or JSONL, at intervals from 20 minutes to weekly. On top of that come exports from the interface and a public API. One limit remains: audit logs, masking at ingestion, retention policies and project-level roles are not part of the MIT core. Anyone relying on those cannot simply fall back to the free self-hosted route.

Regulation and data

Data processing agreementyes, the data processing agreement of its owner ClickHouse, version of 10.04.2026, becomes part of the contractThe former Langfuse address for this document now redirects to the ClickHouse version. The EU standard contractual clauses are incorporated by reference and count as executed with the agreement.Source · As of 2026-07-30
Storage locationEU region Ireland on AWS eu-west-1 under cloud.langfuse.com, selectable at sign-upFurther regions are the United States in Oregon, Japan in Tokyo and a separate region for health data. Data, accounts and infrastructure are fully separated between regions. Switching is not built in, it requires a new account and a migration of your own.Source · As of 2026-07-30
Subprocessorscomplete subprocessor list published, as of 10.04.2026Named are AWS as operator of the customer-selected region, plus Salesforce with Slack, Pylon Labs, Google Workspace, Waypoint Tech, GitHub and kapa.ai for support and operations, all in the United States, along with nine ClickHouse group companies across eight countries.Source · As of 2026-07-30
Third-country transferyes, transfer to the United States does take place, covered by the EU standard contractual clausesThe payload data stays in the chosen region. All support and administrative services, however, sit in the United States. Also relevant: for pay-as-you-go billing the terms name Langfuse GmbH as the contracting party, in every other case ClickHouse. The terms name the law of the US state of Delaware as governing law throughout, with no carve-out for EU customers.Source · As of 2026-07-30
Training on customer datano, expressly ruled outIn its privacy FAQ the vendor states that it does not train or fine-tune ML or LLM models on client data. Model-based evaluations run through the customer's own model connections, so no additional model vendor is introduced by the tool.Source · As of 2026-07-30
Retention and deletionconfigurable per project from a minimum of 3 days up to unlimited, deletion runs nightly; capped by plan: 30 days on the free plan, 90 days on the USD 29 plan, 3 years from USD 199 upwardsDeleting individual traces, whole projects or the entire organisation is possible through the interface and the API. The retention policy itself is one of the licensed features when self-hosting.Source · As of 2026-07-30
CertificationsSOC 2 Type II and ISO 27001 for Langfuse Cloud, audited annually, plus annual independent penetration tests and continuous vulnerability scanningThe public pages do not state clearly at which plan level the report and certificate are actually handed over. That belongs in the conversation before signing.Source · As of 2026-07-30
EU AI Act, Article 50unclearNeither the security page nor the compliance FAQ says anything about the EU AI Act. The tool is not a model provider either, it records calls to other people's models. It can therefore help you meet your own documentation duties, but carries no visible commitment of its own. Ask and get it in writing.no source given · As of 2026-07-30
Audit loggingtamper-evident audit logging only from the USD 2,499 per month plan upwards, when self-hosting only with a paid licence keyThe pricing page lists audit logs and SCIM only in the top plan. When self-hosting, the same features sit behind the enterprise licence key, whose price is not published and which the vendor page ties to a ClickHouse commercial plan.Source · As of 2026-07-30

What it really costs

Entry price

Free for 50,000 units a month, two seats and 30 days of retention. The first paid plan costs 29 dollars a month and includes 100,000 units, unlimited seats and 90 days of retention. Alternatively the MIT-licensed core can be self-hosted at no charge.

Source · As of2026-07-30

What gets expensive

In two places. First consumption: beyond the included quota, each additional 100,000 units costs 8 dollars, falling to 6 dollars at volume. A chain with many intermediate steps per request produces a multiple of the expected units, and nobody models that in advance. Second the governance features: single sign-on and roles add 300 dollars a month, while audit logs and SCIM jump the bill from 199 to 2,499 dollars a month. If you only need those because an audit demands them, you pay that jump mid-year.

What it displaces

  • Model calls with no record at all
  • Prompt versions hard-coded in the codebase
  • Evaluation runs kept in a spreadsheet

Interfaces

  • Ingestion via OpenTelemetry
  • SDKs for Python and JavaScript or TypeScript
  • Public REST API
  • Scheduled export to S3, Google Cloud Storage or Azure Blob Storage in Parquet, CSV, JSON or JSONL
  • Export from the interface as CSV, JSON or JSONL

At Convios: Not used by us

Not used in our own operation. This assessment rests on the public documents of the vendor and its owner, not on hands-on experience.

Matching methods

Evidence

  1. 1.Plans, included units, retention periods, overage tiers, add-on and audit logging only in the top plan · As of 2026-07-30
  2. 2.Certifications, encryption, penetration testing, available regions · As of 2026-07-30
  3. 3.EU region in Ireland on AWS eu-west-1, full separation between regions, switching only via a new account and a migration of your own · As of 2026-07-30
  4. 4.No training on client data, retention configurable per project, data processing agreement available · As of 2026-07-30
  5. 5.Data processing addendum with EU standard contractual clauses incorporated, version dated 10 April 2026 · As of 2026-07-30
  6. 6.Full list of subprocessors and group affiliates, dated 10 April 2026 · As of 2026-07-30
  7. 7.For pay-as-you-go billing the contracting party is Langfuse GmbH; governing law is Delaware · As of 2026-07-30
  8. 8.Acquisition by ClickHouse on 16 January 2026, commitment to continue under the MIT licence and to keep self-hosting available · As of 2026-07-30
  9. 9.Legal entity Langfuse GmbH, Charlottenburg local court HRB 248821B, managing directors Aaron Katz and Clemens Rawert · As of 2026-07-30
  10. 10.MIT core without limits; audit logs, masking, retention policies, project-level roles and SCIM require a licence key · As of 2026-07-30
  11. 11.Scheduled export of traces, observations and scores in Parquet, CSV, JSON or JSONL · As of 2026-07-30
  12. 12.Vendor adoption figures: GitHub stars, Docker pulls, monthly SDK installs, use across Fortune 50 and Fortune 500 companies · As of 2026-07-30

Last reviewed: 2026-07-30byDr. Oliver Gausmann, Convios GmbH

Details out of date? Let us know.

Work through this method with an AI