Langfuse vs. Pydantic Logfire

This guide outlines the key differences between Langfuse and Pydantic Logfire to help engineering teams choose the right AI engineering platform. All Logfire facts were checked against public Pydantic pages in October 2026.

TL;DR

• Open-source core, free to self-host. Langfuse's core tracing, evaluation, prompt management, and APIs are MIT-licensed and available on Cloud or self-hosted; additional Enterprise features require a license. Self-hosting is free on every tier, with Enterprise SSO and organization RBAC included. Logfire's SDK is MIT, but the Logfire platform is commercial, and self-hosting is an Enterprise deployment.
• Works across frameworks and model providers. Langfuse ingests OpenTelemetry and native integrations, Pydantic AI and LiteLLM included, and judges run on your own model connections. Logfire is the Pydantic team's first-party backend; its gateway supports bring-your-own provider credentials with no markup, and built-in providers add 5% on Personal and Team and 3% on Growth.
• 3 years of history and unlimited users on a published plan. Langfuse Pro keeps 3 years with no seat fee; Logfire keeps 30 days on Personal and Team and up to 90 days on Growth, longer retention is an Enterprise conversation, and Team caps seats at 12. For a 10-person team that needs 90 days, Langfuse Core is a quarter of the Logfire Growth price at 500k units a month.
• Data through APIs and exports, or SQL when self-hosting. Langfuse runs on ClickHouse, open source since 2016, with observation and metrics endpoints, scheduled exports to your bucket on Pro with the Teams add-on, Enterprise, and self-hosted, and SQL on your own ClickHouse. Logfire exposes SQL querying with plan-specific daily allowances, and its public API starts on Growth.
• The improvement loop inside the platform. Langfuse runs judges and code evaluators on live observations, prompt and model experiments in the UI, annotation queues on every plan, and prompt versions that carry model and tool config; complete applications run through SDK experiments with results compared in Langfuse. On Logfire, experiments and online evaluators run from your code, annotation queues are Design Partner early access, and a prompt version freezes the template text only.
• Same-day switch. Repoint the OpenTelemetry exporter or the Pydantic AI integration; datasets import via SDK or API. Most teams see first traces in Langfuse the same day.

Choose Langfuse if you want an open-source core and the improvement loop inside the platform: evaluators that run on live traffic, prompt experiments from the UI and application experiments through the SDK, annotation queues on every plan, prompt versions that carry model and tool config, 3 years of history and unlimited users on Pro, on ClickHouse you can query and export. Proven at 21 of the Fortune 50. You accept that host metrics, logs, and service maps stay in the APM you already run.

Choose Logfire if you want application performance and AI traces on one record meter, $0 scores, and a 10-million-record free allowance. You accept a commercial platform with self-hosting on Enterprise, 30 or 90 days of history below Enterprise, seat caps on Team, daily query allowances, experiments and online evaluators that run from your code, and prompt versions that freeze only the template.

---

Open source and deployment

Langfuse's core tracing, evaluation, prompt management, and APIs are MIT-licensed and available on Cloud or self-hosted; additional Enterprise features require a license. Self-hosting runs the UI, login, and data in your environment on every tier, including air-gapped operation, and Enterprise SSO and organization-level RBAC ship in the free distribution. Langfuse Cloud is covered by SOC 2 Type II and ISO 27001 audits, is GDPR compliant, and signs a BAA from the Pro plan in the HIPAA region. Langfuse is used by 21 of the Fortune 50 and 129 of the Fortune 500.

The Logfire SDK is MIT, and the pricing FAQ says the Pydantic library stays MIT. Logfire, the platform, is a commercial product. Enterprise Self-hosted runs an open-sourced Helm chart on your Kubernetes cluster, with Postgres and any S3-compatible backend. Retention on that deployment is customer-determined. SSO, SCIM, and custom roles start on Enterprise.

---

Storage and data access

ClickHouse has been open source since June 15, 2016 (Apache 2.0) and is today the most popular open-source analytical database, with 2,000+ contributors. Langfuse Cloud and self-hosted run the same ClickHouse architecture, so you can move between them without locking into a closed store.

Langfuse works directly with ClickHouse's core team as part of ClickHouse. Every LLM call, tool execution, and agent step is written once to a wide, immutable observations table (details). High-volume reads go through the Observations and Metrics APIs, scheduled exports (Pro with the Teams add-on, Enterprise, and self-hosted) land JSONL or Parquet in your own S3, GCS, or Azure bucket, and self-hosters query ClickHouse in SQL.

Logfire exposes a SQL Workbench in the UI and a SQL query API that returns JSON, Arrow, or CSV. The pricing page meters that API by plan: up to 500 requests a day on Personal and Team and up to 5,000 a day on Growth, and the public API is a Growth feature. Enterprise Self-hosted documents Postgres and S3-compatible storage under the Helm chart.

The two APIs differ in shape. Langfuse exposes observation and metrics endpoints with separate plan-specific limits; Logfire exposes SQL querying with plan-specific daily allowances. One SQL query is not one paginated API request, so compare returned data, pagination, export volume, and the queries your workflow needs rather than request counts.

---

Integrations and framework coverage

Langfuse works across frameworks and model providers, including Pydantic AI and LiteLLM, through native integrations and the OTLP endpoint for any language. Judges run on your own model connections, and routing stays with whichever gateway you already use.

Logfire is the Pydantic team's OpenTelemetry backend and the first-party path for Pydantic AI, with documented integrations for OpenAI, Anthropic, LangChain, LlamaIndex, LiteLLM, CrewAI, and other frameworks on top of any OpenTelemetry instrumentation; its pricing page covers agents, services, and hosts on the same record meter. Its gateway supports bring-your-own provider credentials with no markup, up to 3 on Personal and Team and unlimited on Growth; built-in providers add 5% on Personal and Team and 3% on Growth, and the Enterprise gateway is an add-on.

---

Pricing

Logfire includes 10 million records on Personal, Team, and Growth, then $2 per million on Team and Growth. A record is a span, a log, or a metric. The pricing page lists $0 per 1,000 scores on Team and Growth. Personal hard-caps at 10 million records and pauses ingestion. Team and Growth can set a spending cap.

Langfuse bills units at $8 per 100,000 after the included 100,000, then $7 per 100,000 from 1 million to 10 million. A trace, an observation, and a score each count as one unit. There is no separate seat fee on Core and above.

Worked example

A 10-person team that needs 90 days of history. 50,000 traces a month, each with 8 observations and one trace-level score (a judge verdict or user feedback on the final output). In Langfuse that is 1 trace + 8 observations + 1 score = 10 units per trace, 500,000 units a month. Scoring every observation instead would make it 17 units per trace. In Logfire the 8 observations are 8 spans, 400,000 records; the trace itself is not a record and scores carry no per-score charge. Logs, metrics, and extra spans would add records, and the month stays inside the 10 million allowance either way. Public list prices.

Langfuse usage is 400,000 billable units at $8/100k. Core is a quarter of Logfire Growth, the plan that matches its 90 days and unlimited users, and Pro with 3 years still comes in under Growth. Team is the cheaper Logfire plan but needs five $25 seats for this team, caps at 12, and keeps 30 days.

Where the meters cross. With 90 days on both sides, Langfuse Core stays below Logfire Growth up to roughly 3.1 million units a month, and below a fully seated 12-person Team ($49 + 7 × $25 = $224) up to roughly 2.75 million. Langfuse Pro with 3 years stays below Growth up to roughly 725,000 units. Above those volumes, Logfire's $2 per million record price is the lower bill regardless of retention.

Each total is software subscription plus seats plus telemetry usage. Model inference is a separate invoice on both products: Langfuse judges use your model connection, and Logfire gateway calls carry no markup on your own credentials and 5% or 3% on built-in providers. Optional add-ons stay outside the table: Langfuse's $300/month Teams add-on for scheduled bucket exports and Cloud SSO, and Logfire's Enterprise gateway. A score-heavy loop widens Langfuse's bill, because each score is a unit; Logfire lists scores at $0 per 1,000 on Team and Growth.

See Langfuse pricing or talk to us for a workload that also needs multi-year retention.

---

Evaluation and human review

Logfire's evaluation workspace stores hosted datasets or reads cases from code. Experiments run through Pydantic Evals, a code-first library. The web UI prepares cases and compares a candidate to a baseline. The docs state that the UI does not execute the task under test. Online evaluation follows the same pattern: the Live Evaluations page monitors results that evaluators emit from your application through Pydantic Evals' @evaluate decorator or the OnlineEvaluation capability, and the docs point to the Pydantic AI guide to add an evaluator.

Run annotations record a pass, neutral, or fail on one agent run, with a comment and tags. They are in beta, and the human-review page says they are available in every project. They do not roll into experiment or live-evaluation aggregates. Annotation queues are a Design Partner feature, in an early-access program. The pricing page lists human annotations on Team and above, so the contracted plan is the one to confirm.

Langfuse runs LLM-as-a-judge and code evaluators inside the platform, on live observations and in experiments, with no evaluator code deployed in your application. Prompt and model experiments execute directly in the UI; complete applications and agents run through SDK experiments or your own execution environment, with results compared in Langfuse. Annotation queues are part of the shipped product; Pro includes unlimited queues. Score analytics compares judge scores with human labels.

---

Prompt management

Both products version prompt text and promote a label such as production.

Logfire's prompt UI authors a template, tests it against scenarios, and saves a version. The docs are specific about the boundary: a version freezes the template text. It does not freeze scenario messages, model settings, tool definitions, managed-variable fragments, or rollout rules. Promotion happens on the backing managed variable. Tool definitions in prompt settings affect runs in the UI; the application still configures its own tools.

Langfuse stores model, temperature, tool definitions, and JSON schemas on the prompt version, promotes that version with labels, and can reference other prompts so shared instructions live in one place. The running app fetches the labeled version and can link it to the trace.

The difference shows when a change touches more than text. When you change model, temperature, tool definitions, and prompt text together, Langfuse can version that configuration with the prompt, so a rollback restores all of it; Logfire's prompt version freezes the template text, and the other settings are managed separately. In both products the application must read and apply what it fetches: storing tool configuration on a Langfuse prompt does not change application behavior until your code uses that config.

---

Application context

Logfire's pricing page covers agents, services, and hosts: Kubernetes, databases, logs, and service maps on the same record meter. That is the reason a platform team picks one tool for the request path and the agent.

Langfuse records the AI trace: generations, tools, retrieval, sessions, agent graphs, and cost. Host metrics and APM stay in the observability platform you already run. The two products can sit side by side: Logfire for the service, Langfuse for the evaluation and prompt workflow. Pydantic AI emits OpenTelemetry that any OTel backend, Langfuse included, accepts.

---

Which should you choose

Choose Langfuse if you want an open-source platform that works across your stack: OpenTelemetry instrumentation for any framework, free self-hosting with SSO and organization RBAC included on every tier, judges and code evaluators that run inside the platform, prompt experiments from the UI and application experiments through the SDK, annotation queues and prompt versions that carry model and tool config, data in ClickHouse you can query and export, and 3 years of history with unlimited users on Pro.

Choose Logfire if the decision is one OpenTelemetry backend for services and agents: a low record price, $0 scores, and host, database, and service telemetry on the same meter, with up to 90 days of history on Growth.

Run both when Logfire remains the application monitor and Langfuse holds datasets, reviews, and prompt releases.

---

Why teams move away from Logfire

The same reasons recur in migration conversations.

• Self-hosting routes through Enterprise. The Logfire platform is commercial, and the open-sourced artifact is a Helm chart that ships with the Enterprise Self-hosted deployment. Langfuse self-hosts its open-source core, UI, login, and data, on every tier, with SSO and organization RBAC in the free distribution, and the same core runs on Cloud when teams move later.
• History below Enterprise stops at 90 days. Logfire Growth retains data for up to 90 days; Personal and Team retain 30, and unlimited retention is an Enterprise feature per the pricing FAQ. Langfuse Pro includes 3 years on the published plan, so the failure from last quarter is still there to replay or turn into a dataset item.
• Experiments and online evaluators run in your code. Logfire's UI prepares datasets and reads results; the docs say it does not execute the task, and Live Evaluations monitors evaluators that your application runs through Pydantic Evals. Langfuse runs judges and code evaluators inside the platform on live observations, runs prompt experiments from the UI and full applications through the SDK, compares rows to a baseline, and gates CI on the result.
• A prompt version is the template. Model settings and tool definitions are not frozen with the Logfire version. Langfuse versions that config with the text and promotes it with a label, so a rollback restores the whole configuration once the application re-reads it.
• Annotation queues are an early-access program. Logfire's direct run annotations are in beta, and queues are Design Partner early access. Langfuse annotation queues ship in the product, unlimited on Pro.
• Programmatic access differs in shape and allowance. Logfire exposes SQL querying with 500 requests a day on Personal and Team and 5,000 on Growth, and the public API starts on Growth. Langfuse exposes observation and metrics endpoints with their own plan-specific limits, scheduled exports on Pro with the Teams add-on, Enterprise, and self-hosted, and SQL for self-hosters. Compare the data returned and the queries you need, not request counts alone.

---

Why teams choose Langfuse

A pattern repeats across production adopters: start where governance requires, often self-hosted, then grow into the platform.

• Canva instruments Java through OpenTelemetry and Python through the Langfuse SDK, then scores both with LLM-as-a-judge evaluators across 15–20 metrics. Non-engineers build evaluators. They self-hosted first, then moved to Cloud.
• Merck self-hosts about 80 GenAI use cases for 200+ people, provisioned via API, for data sovereignty.
• Khan Academy built a custom Go client on the open API, with 100+ users across 7 product and 4 infrastructure teams.

---

Switching from Logfire

Repoint telemetry and recreate the assets the other system will own. History stays in Logfire for that plan's retention window. Most teams see first traces in Langfuse the same day.

• Instrumentation: point the OpenTelemetry exporter at Langfuse, or use the Pydantic AI integration. Both backends can receive spans during the overlap.
• Datasets and prompts: fetch cases with the Logfire Datasets SDK and import them through the Langfuse SDK or API. Recreate prompt templates in prompt management, put model and tool settings into the version config if they lived outside the Logfire template, and make sure the application reads and applies that config.
• Evaluators: Pydantic Evals stays in your repository for code-defined experiments. Langfuse code evaluators and judges are the in-platform equivalents for online scoring and prompt experiments, so the @evaluate and OnlineEvaluation hooks can come out of the application.

Talk to us if you want help planning the overlap.

---

Start free: Cloud or self-host

Langfuse Cloud Hobby includes 50k units/month, no credit card. Or self-host the MIT stack. The example project is populated before you instrument anything.

---

FAQ

---

---