Langfuse Pricing: Cloud Plans, Self-Hosting, Total Cost, and Ownership

Langfuse pricing has two ledgers. Langfuse Cloud charges a subscription plus usage after the included allowance. A self-hosted deployment shifts the service onto infrastructure and people your organization operates. Comparing the Cloud card with a monthly compute estimate misses retention and support as well as ingestion peaks and database operations. Backup cost disappears from that estimate too. It also omits upgrades and the enterprise controls your team expects.

The public numbers below were reviewed on October 26, 2026. Treat them as a procurement snapshot and recheck the current Langfuse pricing page before signing an order.

TL;DR

• Langfuse Cloud lists Hobby at free, Core at $29 per month, Pro at $199 per month, and Enterprise at $2,499 per month before optional add-ons or custom volume terms.
• The public plans package included usage units and historical data access, with throughput and support controls. Security controls, commercial terms, and plan limits complete the purchase.
• Self-hosting requires two application containers plus PostgreSQL and ClickHouse, with Redis or Valkey and S3-compatible object storage.
• Compare twelve-month cost using measured ingestion and retained data, then support work and upgrades. Add backups, on-call ownership, and recovery tests.

Langfuse pricing packages several different meters

The Cloud plan cards publish a base subscription and an included monthly unit allowance. Hobby includes 50,000 units per month. Core and Pro each list 100,000 included units. Enterprise lists the same allowance, while Core and Pro show additional usage at $8 per 100,000 units and lower rates at volume. Enterprise also shows the same public overage reference alongside custom volume pricing.

Keep the word unit exactly as Langfuse uses it. A unit is the commercial meter on the pricing page. It should never be silently converted into model tokens or traces, and it is separate from dollars of provider spend. Measure a representative application's billed units in Langfuse, then apply the pricing calculator or a written quote. Provider inference charges remain a separate invoice.

A useful workbook gives base subscription, included units, measured units, public or quoted overage, and model-provider spend separate rows. That prevents a low Langfuse platform bill from being mistaken for the full LLM operating cost. LLM observability tools explains why tracing products also differ in their capture model, which affects the volume they ingest.

Cloud plan entitlements change the buying decision

Hobby is free and includes 50,000 units per month plus 30 days of data access. It allows two users and provides community support through GitHub. Langfuse presents it for hobby projects and proofs of concept. The two-user limit and short history make it a poor production baseline for a security review, even if the measured ingestion fits.

Core lists $29 per month. It includes 100,000 units and 90 days of data access. Users are unlimited, with support available in the application. The comparison surface also lists higher ingestion and API limits than Hobby. Core is the first plan where a small production team can separate users without buying seats, but its 90-day history still needs to match incident and audit timelines.

Pro lists $199 per month. It extends historical data access to three years and adds data-retention management plus unlimited annotation queues. The plan also raises rate limits and prioritizes support in the application. It includes security or compliance capabilities shown on the pricing page. An optional Teams add-on is listed at $300 per month and includes enterprise SSO, SSO enforcement, fine-grained RBAC, and a dedicated Slack or Microsoft Teams support channel.

Enterprise lists $2,499 per month and routes the buyer to sales. The page adds audit logs, SCIM, custom rate limits, uptime and support SLAs, a named lead support engineer, optional yearly commitment, architecture reviews, invoice or AWS Marketplace billing, and vendor-onboarding support. Verify every required entitlement in the order form. A pricing-page checkmark is discovery evidence, while the signed agreement controls the purchase.

I think a team choosing Hobby for a production service because the unit allowance fits has optimized the smallest line in the budget. Thirty days of accessible history and community support will matter long before the saved subscription does.

Retention changes both price and evidence

Langfuse labels the plan field Access to historical data. Hobby shows 30 days and Core shows 90 days. The Pro card shows three years. Enterprise inherits the Pro package on the public card, subject to the final agreement. Data-retention management appears on Pro and above.

Procurement should distinguish accessible history and configured retention. Export, deletion, backup retention, and restoration can each follow another clock. Ask how a plan change affects older records. Also ask how long scheduled exports remain retrievable and which data is included in retention controls. The DeepInspect versus Langfuse comparison separates application trace history from an identity-bound policy decision record. They may share a request ID, but they answer different review questions.

Run the retention test with a record at the edge of the required period. The evidence reviewer should be able to open the trace, follow its observations, identify its prompt version and scores, and export the required fields. Put the date and timezone on the test sheet. A screenshot of a dropdown proves a setting existed; it proves nothing about retrieval three years later.

Self-hosting moves cost into your infrastructure

Langfuse's self-hosting documentation says the software can run through Docker on an organization's infrastructure and notes that some add-on capabilities require a license key. Docker Compose is positioned for local use and testing on a single virtual machine. High availability and scaling need a production deployment path, along with backups and recovery. Production guidance points to Kubernetes with Helm or cloud-specific deployment paths, with the organization responsible for the infrastructure.

The published architecture has two application containers. Langfuse Web is the user interface and API service. Langfuse Worker processes events asynchronously. Those services depend on PostgreSQL for transactional data and ClickHouse for traces and scores. Redis or Valkey handles cache and queues, while S3-compatible blob storage retains events and attachments. Playground and evaluation features may also call an external LLM API or gateway.

Draw these components as boxes before pricing them. On the left, place the web and worker deployments. On the right, stack PostgreSQL and ClickHouse above Redis or Valkey and object storage, then draw backup and monitoring paths underneath. The diagram turns one line called self-hosted Langfuse into the services an on-call engineer must keep healthy.

The self-hosted TCO ledger

Start with infrastructure consumed under ordinary load and at the highest observed ingestion burst. Price application replicas, database compute, ClickHouse storage, object storage, cache memory, network transfer, load balancing, and secrets. Add certificates and log collection beside the backup copies. Add a recovery environment if the service has a recovery-time objective.

Labor belongs on the next set of rows. Someone owns initial deployment, upgrades, schema migrations, and capacity planning. Add security patches, backup tests, restoration drills, alert tuning, and incident response. The self-hosting page describes queued ingestion that writes incoming events to object storage before workers ingest them into ClickHouse. That design handles bursts, but it also gives operators a queue and object store beside the worker fleet and analytical database they observe.

Use twelve months of expected cost across normal and forecast workloads, then add a separate burst case. Three cases are useful here because infrastructure gets sized for a peak while usage fees follow actual consumption. Record engineering hours with the loaded internal rate. Include any paid self-hosted entitlement and support quote as its own line.

The resulting number is an operating estimate, rather than a promise. A tiny development deployment may cost less than Cloud when staff already operate the stack. A production service with high availability, formal backup tests, a support SLA, and SSO can reverse that result quickly. Self-hosted AI gateway planning covers similar control-owner questions for a different service category.

Support is part of the production architecture

Cloud support progresses from GitHub community support on Hobby to in-app support on Core, prioritized in-app support on Pro, and named support plus SLAs on Enterprise. The Teams add-on introduces a private channel. Capture the response-time objective, coverage hours, escalation route, severity definitions, and architecture-review entitlement in the purchase record.

Self-hosted buyers need two support assignments. The vendor or community can help with Langfuse application behavior. Your platform team still owns cluster health, network policy, database recovery, object storage, and upgrades. A support contract cannot repair an unavailable customer-operated ClickHouse cluster unless the contract explicitly includes that work.

Test support before the renewal decision. File a representative technical case and record the first useful response. Then see whether the responder can work from the diagnostics your team is permitted to share. That produces better evidence than comparing support labels on a card.

A Cloud versus self-hosted acceptance test

Take one production-shaped trace population and ingest it into the candidate deployment for a full billing interval where practical. Record billed units, peak requests per minute, database growth, object-storage growth, query latency, and operator time. Then restore a backup into a clean environment and retrieve a named trace.

For Cloud, confirm the plan exposes the required history, access controls, audit capabilities, and throughput. Verify the export path and support route too. For self-hosted, confirm high availability, queue recovery, database backup, and object restoration. Verify the upgrade procedure and monitoring separately. Keep model-provider cost outside both columns.

The decision should be reversible. Preserve export procedures and stable trace identifiers. If a future move changes identifiers, keep a mapping that supports incident review. The Langfuse alternatives guide covers product-category choices; this test answers the narrower financial and operational question.

DeepInspect

Langfuse captures engineering traces and supports prompt and evaluation workflows. Its product also supports observability operations. DeepInspect evaluates policy on authenticated HTTP model traffic routed through it. The two systems can coexist by carrying one correlation identifier while keeping the trace and policy-decision records separate.

That boundary matters in the cost model. Langfuse infrastructure, units, retention, and support belong in the observability ledger. DeepInspect belongs in the enforcement and identity-bound evidence ledger. Combining the two into a single AI platform amount makes renewal analysis harder and assigns outages to the wrong operator.

Book a technical deep dive at deepinspect.ai.