Langfuse+46 Positive

GILD Review

GILD's call and the buying facts behind it.

From GILD research

Should we adopt Langfuse?

Which costs your team less, self-hosting Langfuse with its ClickHouse stack or paying Cloud's per-unit bill?

30-Second TakePick Langfuse for tracing, evals and prompt versioning when trace data must stay in your own infrastructure.
Best fitTeams that want tracing, evaluation, and prompt management from one open-source platform, above all those that must keep trace data in their own infrastructure.GILD Review
Consider another workflowFor LangGraph agents with vendor hosting, evaluate LangSmith. For managed evals as the main workflow, evaluate Braintrust. If your team also monitors classic ML models, evaluate Arize AX.GILD Review

Drawn from 103 sources and 17 operator insights

GILD analysis

Langfuse Review

Pick Langfuse for tracing, evals, and prompt versioning on one MIT-licensed platform, in Langfuse Cloud or on your own servers; Pinterest, Cresta, SuperDial and Khan Academy trace production agents in it. Bayes Impact and one self-hoster moved to Arize Phoenix for a lighter stack, so plan for ClickHouse, Redis and S3 if you self-host. On Cloud, expect unit billing that grows with trace depth and a v4 cutover on November 16, 2026 that old SDKs will not survive.

AI Observability
AI Costs

The decision

Best fit: Teams that want tracing, evaluation, and prompt management from one open-source platform, above all those that must keep trace data in their own infrastructure.
Consider another workflow: For LangGraph agents with vendor hosting, evaluate LangSmith. For managed evals as the main workflow, evaluate Braintrust. If your team also monitors classic ML models, evaluate Arize AX.
Adoption condition: Use v4 SDKs or OpenTelemetry and mask sensitive fields. Set spend alerts or a retention policy. If self-hosting, assign a ClickHouse owner.

Basis: Langfuse documentation, plans, and pricing, plus 88 public accounts. They are 33 GitHub issues and pull requests, 22 Hacker News comments, 15 Reddit and dev.to posts, 5 X posts, 6 company and personal blog posts, 6 recorded talks, and 1 Product Hunt review. Fifteen records added on Oct 28, 2026 come from research that looked for problems first, so they take no side in the scores. In a pilot, run a week of peak traffic through self-hosted Langfuse and record ClickHouse memory, CPU, and disk growth.

Buying & deployment factsWhat a team building on it should verify before production traffic grows.

How you buy it

US$29

A month for Core with 100,000 units and unlimited users. Hobby is free.

Source 1, Langfuse, Cloud pricing: Cloud pricing Source 2, Langfuse, Self-hosted pricing: Self-hosted pricing

Details

Hobby is free with 50,000 units a month, 30 days of data access, and 2 users. Core is US$29 a month with 100,000 units and unlimited users, Pro is US$199 a month with 3 years of data access, and Enterprise is US$2,499 a month. Self-hosting the MIT-licensed core is free, and self-hosted Enterprise is custom-priced and bundled with a ClickHouse plan.

Cost drivers

US$8 per 100,000

Units past the allowance. Every trace, observation, and score counts as one.

Source 5, Langfuse, Billable units: Billable units Source 1, Langfuse, Cloud pricing: Cloud pricing

Details

Every trace, observation, and score is a unit, including scores from Langfuse's own evaluators. Past the included units, paid plans charge US$8 per 100,000. Rates fall to US$7 above 1 million, US$6.50 above 10 million, and US$6 above 50 million. Pro's enterprise SSO and project RBAC require the US$300 monthly Teams add-on.

Where it runs

US, EU, Japan

Separate cloud regions, plus HIPAA. Moving region means a new account.

Source 3, Langfuse, Data regions: Data regions Source 6, Langfuse, Self-host Langfuse: Self-host Langfuse

Details

Cloud runs on AWS and ClickHouse Cloud in separate US, EU, Japan, and HIPAA regions. Moving regions requires a new account. Use Docker Compose for self-hosted testing. Production uses Kubernetes (Helm) or the AWS, Azure, and GCP deployment guides, with PostgreSQL, ClickHouse, Redis, and S3.

What to verify

November 16, 2026

Cloud goes v4-only. Python SDKs below 4.7.0 and JS SDKs below 5.4.0 must upgrade.

Source 4, Langfuse, Langfuse v4: Langfuse v4 Source 8, langfuse.com, Upgrade to Langfuse v4 (FAQ): Upgrade to Langfuse v4 (FAQ)Source 3, Langfuse, Data regions: Data regions Source 7, Langfuse, Data retention: Data retention

Details

Cloud becomes v4-only on November 16, 2026: legacy ingestion and read APIs stop, Python SDKs below 4.7.0 and JS/TS SDKs below 5.4.0 must be upgraded, and trace-level LLM judges move to observations. Projects the Migration Assistant flags must act before then. Cloud does not replicate ClickHouse traces to another region. Retention policies require Pro or a self-hosted Enterprise key, with a 3-day minimum; without a retention policy, Langfuse does not automatically delete event data. Audit logs come with Enterprise on Cloud and with a self-hosted Enterprise key, and project retention does not delete them.