The short answer
Langfuse, now owned by ClickHouse, combines tracing, evaluations and prompt management in a platform you can self-host or use in the cloud. Arize Phoenix offers a local and self-hosted workflow under the source-available Elastic License 2.0; Arize AX is a separate commercial platform. Comet’s Apache-2.0-licensed Opik offers tracing, evaluation experiments and a prompt library, with self-hosted and hosted options. Start with who will operate the service, then compare licenses, instrumentation and cloud limits.
| Decision | Langfuse | Arize Phoenix | Opik |
|---|---|---|---|
| What it is | Platform for tracing, evaluating and managing prompts for LLM apps and agents | Local and self-hosted AI observability, evaluation and experimentation platform | Open-source AI observability and evaluation platform with tracing and agent testing |
| Owner / company | ClickHouse; acquired Langfuse in January 2026 | Arize AI | Comet ML, Inc. |
| License | MIT core; ee/, web/src/ee/, worker/src/ee/ and packages/shared/src/server/ee/ use the separate ee/LICENSE enterprise terms | Elastic License 2.0: source-available; restricts offering a substantial set of its features as a hosted or managed service to third parties | Apache License 2.0 for the public repository; hosted Enterprise offerings are separate |
| Self-hosting options | Docker Compose for local use and testing; Kubernetes via Helm for production; web and worker containers | CLI, Docker / Compose or Kubernetes / Helm; SQLite by default, PostgreSQL for production | Git and Docker / Compose for local setup; Kubernetes / Helm recommended for production; local Compose is not intended for production |
| Storage to operate | PostgreSQL, ClickHouse, Redis / Valkey and S3 / blob storage | SQLite for local or single-user use; PostgreSQL 14 or newer for production | ClickHouse, MySQL, Redis and MinIO / S3-compatible object storage; Kubernetes ClickHouse uses ZooKeeper for coordination |
| Hosted free tier | Hobby: free, 50k units/month and 2 users | Current Phoenix site directs cloud signups to Arize AX. AX Free: 25k spans and 1 GB ingestion/month; unlimited users | Free Cloud: $0, 25k spans/month and up to 10 team members on the plan card |
| Paid cloud plans | Core $29/month; Pro $199/month; Enterprise $2,499/month; yearly commitment is optional. Each includes 100k units/month; extra usage starts at $8/100k units and falls with volume. Core: 90 days data access; Pro and Enterprise: 3 years | Separate Arize AX plans: Pro $50/month with 50k spans, 10 GB ingestion/month and 30-day retention; Enterprise custom pricing and limits | Pro Cloud $19/month: 100k spans/month, up to 50 team members and 60-day retention. Enterprise: custom pricing and usage |
| Tracing / instrumentation | OpenTelemetry-based Python and JS/TS SDKs; OpenAI wrappers, LangChain integration and direct OpenTelemetry ingestion | OpenTelemetry / OpenInference instrumentation; Python and TypeScript tooling, with OpenAI, LangChain and LlamaIndex integrations | Python and TypeScript SDKs; @track decorator and provider/framework integrations; OpenTelemetry ingestion currently uses HTTP transport |
| Evaluations | LLM-as-judge, code evaluators and human annotation; datasets and experiments via UI or SDK; online and offline workflows | Code-based and LLM-as-judge evaluators; traces, datasets and experiment results; client SDKs or server-side UI evaluators | Heuristic and LLM-as-judge metrics; evaluate application tasks against datasets in experiments |
| Prompt management | Central prompt storage, versioning and retrieval; deployment labels, SDK caching and links to traces | Create, store, load and version prompts; share prompts for collaboration | Prompt Library automatically versions changes, fetches prompts at runtime and links the version used to its trace |
| Free cloud data retention / access | Hobby: 30 days data access | Arize AX Free: 15-day retention. Self-hosted Phoenix retention policies are operated by your team | Free Cloud: 60-day data retention |
Who should pick which
Consider Langfuse when you want tracing, prompt deployment and evaluation workflows in one project, with a choice between hosted cloud and your own infrastructure. ClickHouse acquired Langfuse in January 2026. The official announcement says Langfuse remains self-hostable and that its cloud service continues with the same endpoints. Its core uses MIT; the enterprise directories have separate commercial terms.
Consider Phoenix when you want to run a local workflow backed by SQLite, then operate a shared deployment with PostgreSQL. Its repository documents OpenTelemetry and OpenInference instrumentation, evaluation, datasets, experiments and prompt management. Check the Elastic License 2.0 restrictions before choosing it for a service offered to third parties. Arize says Phoenix has no closed core or commercial tier within Phoenix; AX is a separate platform with its own plans.
Consider Opik when you want Apache-2.0-licensed tracing, dataset evaluation experiments and a prompt library, and can use either Comet’s cloud or your own deployment. Its self-hosting overview says tracing and evaluation are available without the hosted user management features. These are choices based on documented features and operating requirements, not a ranking of output quality or speed.
How to try each in an afternoon
Use the same small application and a few representative inputs for each trial. First inspect one trace, then save a prompt and run an evaluation experiment on a small dataset. Treat this as a proposed workflow; this comparison is based on official documentation checked on November 7, 2026, without hands-on testing.
Langfuse: create a free Hobby account and project API credentials, then follow the tracing quickstart for your existing SDK or framework. For a Python OpenAI application, install langfuse and use its OpenAI wrapper; configure the project credentials and base URL as the guide shows. Inspect the first trace, create a managed prompt, then use the evaluation docs to build a dataset and compare a prompt change in an experiment. Use Docker Compose if you want a local trial.
Phoenix: start locally with uvx arize-phoenix serve, as shown on the current homepage. Choose the OpenInference integration for your framework and send OpenTelemetry traces to that instance. Save and version a prompt, create a small dataset, and run an experiment with a code evaluator or LLM judge. This keeps the trial within Phoenix; a hosted Arize AX trial follows AX’s own documentation and account setup.
Opik: choose Free Cloud or clone the repository and run ./opik.sh for a local Docker deployment. Follow the quickstart: install opik, configure the SDK, and wrap a function with @track or select the integration for your provider. For local use, set OPIK_BASE_URL to http://localhost:5173/api as the self-hosting guide shows. Add a prompt to the Prompt Library, then evaluate your application task against a dataset with a heuristic metric or LLM judge.
Check licenses, limits and operating costs
Cloud allowances use different units. Langfuse defines units as traces plus observations plus scores; Opik and Arize AX list spans, and AX also caps ingestion volume. A single application run can produce several spans or observations. Compare the telemetry your application would send, rather than treating every allowance as a count of user requests. Langfuse labels Hobby’s historical limit as data access; that label alone does not establish a deletion timetable.
Self-hosting means operating the storage and application services listed above, including persistence, backups and upgrades. Langfuse and Opik document several storage services; Phoenix can start with SQLite and recommends PostgreSQL for production. Both Langfuse and Opik position their local Compose setup for testing rather than production. Plan infrastructure and any model calls used by LLM judges alongside the platform subscription.
Use the current product and plan pages when checking hosted options. Phoenix’s current homepage directs cloud signups to Arize AX, so the AX prices and retention figures here do not describe a paid Phoenix edition. Opik’s pricing cards list up to 10 Free Cloud members and 50 Pro Cloud members, while the FAQ on that same page says all plans include unlimited team members. Confirm the member allowance before subscribing. Prices and limits can change; recheck them before committing a workload.
Sources & method
Official documentation was checked on 2026-11-07. Selection advice is our assessment. Interfaces and requirements may change; check the linked source for the version you intend to use.
- Langfuse company facts and ownership
- Langfuse joins ClickHouse: official acquisition notice
- Langfuse root license and enterprise directory exceptions
- Langfuse enterprise license
- Langfuse self-hosting and architecture
- Langfuse cloud pricing and plan limits
- Langfuse tracing quickstart
- Langfuse evaluation overview
- Langfuse prompt management
- Langfuse clarifications: SDKs and billable units
- Arize Phoenix repository and OpenInference instrumentation
- Arize Phoenix homepage and current hosted signup path
- Arize Phoenix Elastic License 2.0
- Phoenix self-hosting options
- Phoenix architecture and database requirements
- Arize Phoenix and AX: separate products and operating models
- Arize AX pricing and retention
- Phoenix tracing and integrations
- Phoenix evaluation methods
- Phoenix prompt management and versioning
- Opik product overview
- Opik Apache License 2.0
- Opik self-hosting overview and feature scope
- Opik local Docker Compose deployment
- Opik platform architecture and storage services
- Comet Opik cloud pricing, limits and retention
- Opik quickstart and instrumentation
- Opik OpenTelemetry transport and SDK support
- Opik dataset evaluation and experiments
- Opik Prompt Library: versions and runtime retrieval
How SOTA reviews projects · Suggest a correction · Download the evaluation checklist
