Langfuse vs. Arize AX / Arize Phoenix

This guide compares Langfuse, Arize Phoenix, and Arize AX. All Arize facts were checked against public Arize sources in September 2026.

Phoenix and AX are different products. Phoenix is source-available (Elastic License 2.0) on SQLite or PostgreSQL. AX is proprietary on adb ( public launch June 2025). Langfuse is one MIT-licensed codebase on ClickHouse ( open source since 2016), identical on Cloud and self-hosted.

On August 13, 2026, Dynatrace announced a definitive agreement to acquire Arize. The transaction has not closed; this page compares AX and Phoenix as they ship today.

TL;DR

Choose Langfuse if you want:

1. The same product in production. Tracing, evaluations, prompt management, and experiments in the MIT distribution. Cloud and self-host share the codebase. Used by 21 of the Fortune 50.

2. An open data plane. Every span lands in ClickHouse, open source since 2016 with 2,000+ contributors. Self-hosters query it in SQL. Cloud and self-host share the same engine.

3. One meter, years of history, code evals on every plan. A unit is a trace, an observation, or a score. Pro includes 3-year retention. Custom code evaluators are not an Enterprise gate.

Choose Arize Phoenix if you want one container on SQLite (Postgres when you outgrow it), OpenInference as the native instrumentor, and client-side evaluators as code you run yourself. You accept Elastic License 2.0 and a different product when you need managed production.

Choose Arize AX if you already run classical ML on Arize, or you want Signal, Alyx, and session-level LLM judges on a managed platform. You accept span + ingested-GB billing, 30-day Pro retention, and custom code evaluators on Enterprise.

Open source and distribution

Langfuse is MIT-licensed across tracing, evaluations, prompt management, and experiments. Self-hosting runs the same codebase and ClickHouse architecture as Langfuse Cloud, so you can move between Cloud and self-host without rewriting the application layer. Enterprise governance modules, including audit logs, project-level RBAC, SCIM, protected labels, and retention policies, need an Enterprise license when self-hosted.

Arize ships two products. Phoenix is source-available under the Elastic License 2.0 (not an OSI-approved open-source license) and self-hosts on SQLite or PostgreSQL. Arize AX is proprietary; self-hosting is an Enterprise option. Phoenix and AX do not share a storage engine: Phoenix uses Postgres/SQLite, while AX runs on adb, Arize's proprietary OLAP database ( public launch June 25, 2025).

Dynatrace's acquisition announcement describes Arize as "OSS-native" and highlights its open-source community, but it does not name Phoenix or state a license or continuity commitment for a standalone open-source product.

History and data plane

Phoenix starts as one container on SQLite. A production Langfuse self-host runs ClickHouse, Postgres, Redis, and blob storage. Docker Compose brings that stack up on one VM. Helm and Terraform are the high-availability paths. Agent traces are deep trees with large payloads and high-cardinality group-bys, so Langfuse left Postgres in v3. How Langfuse runs ClickHouse at agent scale is the design write-up. Joining ClickHouse in January 2026 put the team that builds that engine and the team that runs it on Langfuse Cloud in one company. The public commitment keeps the MIT license, self-hosting, Cloud endpoints, and the same roadmap.

High-volume reads use the Observations and Metrics APIs; self-hosters can query ClickHouse in SQL.

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.

The Arize pricing page lists 15 days of retention on Free and 30 days on Pro; Enterprise retention is custom. AX Cloud stores telemetry in adb, a proprietary store with a June 2025 public launch. Warehouse SQL is Data Fabric: Enterprise-only and currently waitlisted.

Pricing

Langfuse Cloud bills units: one unit is a trace, an observation, or a score, with no seat fees. Arize AX bills on two meters: span volume and ingested GB. On AX Pro, spans beyond the included 50,000 cost $0.0008 each and ingestion beyond 10 GB costs $3/GB, shown as tooltips on the Arize pricing page. Phoenix self-hosted has no platform usage fee; you run the infrastructure. The editable pricing model takes your volumes as inputs so you can find your own crossover.

Worked example

500,000 traces/month · 10 spans (observations) per trace · 1 score per trace · 5 KB average span size · 5 users.

Public list prices; edit the inputs yourself in the pricing model. Langfuse: 6,000,000 units (traces + observations + scores), 100k included → $199 base + $72 (100k–1M at $8/100k) + $350 (1M–6M at $7/100k) = $621, 3-year history included. Arize AX: 5,000,000 spans and 25 GB → $50 base + 4,950,000 × $0.0008 ($3,960) + 15 GB × $3 ($45) = $4,055. Arize does not publish a price for retention beyond AX Pro's 30 days, so none is included.

Tracing and integrations

Both platforms ingest OpenTelemetry. Langfuse maps OpenInference attributes natively, so existing OpenInference instrumentors can keep running and export to Langfuse via OTLP.

Langfuse records a trace per request with typed observations for LLM calls, tools, and retrieval. Multi-turn chats group into sessions, agent runs render as agent graphs, and cost is attributed per generation. Arize AX offers agent trajectory visualizations, session support, and multi-modal tracing on all plans, plus the Alyx assistant and per-plan Signal issue quotas (10/month Free, 25/month Pro, unlimited on Enterprise).

Evaluation and experiments

Langfuse evaluations include LLM-as-a-judge evaluators, code evaluators on live observations and experiments, custom scores via SDK/API, annotation queues, and dataset experiments, Cloud and self-hosted. CI/CD gates fail a pipeline on experiment results. Coding agents can run the same loop through the CLI, MCP server, and agent skill.

Arize AX includes unlimited online and offline evaluations, human annotations, and labeling queues on all plans. The Arize pricing page lists agent-as-a-judge on all plans, plus session- and trajectory-level judges. Langfuse LLM-as-a-Judge targets observations (recommended) or traces (deprecated). For a multi-turn conversation, evaluate the observation that holds the history. The pattern is in evaluating sessions. Session scores and session annotation queues still apply.

Custom code evaluators are Enterprise-only on Arize AX. Langfuse code evaluators and CI gates run on every plan.

Prompt management

Edit the prompt in the UI, test it in the playground, and the running app fetches it with version control, client-side caching, and composability. In Langfuse, pass the fetched prompt into the generation to link it to traces, so each version gets cost, latency, and score metrics. Arize AX includes prompt serving and versioning, playgrounds, multi-prompt comparison, and Prompt Learning optimization on all plans. Phoenix also ships prompt management.

Enterprise security and governance

Langfuse lists a GDPR DPA on all Cloud plans, and SOC 2 Type II and ISO 27001 plus a HIPAA-ready region on Pro or higher. Arize lists SOC 2 Type II, HIPAA, PCI DSS 4.0, and CSA Star Level 1; on the AX plan table, SOC 2 Type II and GDPR start at Pro while HIPAA, Enterprise SSO, and audit logs are Enterprise-only.

Which should you choose

Langfuse if you want one MIT product on an open ClickHouse store. That includes the AI engineering loop, code evaluators and CI experiments, and years of queryable history, on Cloud or self-host.

Arize Phoenix if you want a single container on SQLite or Postgres, OpenInference as the native instrumentor, and client-side evaluators you run yourself.

Arize AX if one managed platform must cover classical ML and GenAI, or you want Signal, Alyx, and session-level judges and will take span+GB billing with 30-day Pro retention.

Why teams move away from Arize AX

• Every span and every GB bills. Arize AX meters span volume and ingestion together; at 500k traces/month the worked example lands at $4,055 on AX Pro vs. $621 on Langfuse Pro, which meters units only.
• Parts of the quality loop are Enterprise-gated. Custom code evaluators are Enterprise-only on Arize AX. Langfuse runs code evaluators and CI experiments on every plan and self-hosted.
• Self-hosting below Enterprise means a different product. Arize AX self-hosting is Enterprise-only; the free path is Phoenix, a separate product on a different storage engine under the Elastic License. Langfuse self-hosts the same codebase as Cloud, MIT-licensed.
• AX Cloud uses a closed store that launched in 2025. adb is proprietary (public launch June 2025). Langfuse runs on ClickHouse, open source since 2016, with the same architecture on Cloud and self-hosted.
• 30 days of history below Enterprise. Arize AX includes 15 days on Free and 30 on Pro, with no public price for more. Langfuse Pro includes 3-year retention.

Why teams choose Langfuse

• Canva built a multi-agent support system for a product with 250 million monthly active users. Help Assistant is Java via OpenTelemetry; Omni Agent is Python via the Langfuse SDK. LLM-as-a-judge evaluators score both systems across 15–20 metrics, and domain experts build evaluators without engineering help. They self-hosted first, then moved to Cloud.
• SumUp rolled AI support to 35+ markets and deflected almost 50% of conversations, cutting external BPO cost by 30%. They started self-hosted for the PoC, then moved to Langfuse Cloud.
• Merck runs about 80 of 300+ GenAI use cases on Langfuse with 200+ people building on it. They self-host for data sovereignty and provision via API.

Switching from Arize

Move live instrumentation and durable assets. Trace history usually stays behind for its retention window. Most teams see first traces in Langfuse the same day. Existing OpenInference instrumentation keeps working because Langfuse maps OpenInference attributes over OTLP.

• From Arize AX: follow the AX migration guide.
• From Phoenix: follow the Phoenix migration guide.
• Datasets, prompts, and evaluators: import datasets via CSV, SDK, or API; recreate prompts in prompt management and judges as evaluators.

Start free: Cloud or self-host

Langfuse Cloud Hobby includes 50k units/month, no credit card. Or self-host the MIT stack. Explore the example project first. Already on Arize? See Switching from Arize.

Get started free
Talk to us

FAQ

Is Langfuse an alternative to Arize AI?

Yes. Langfuse is an open-source alternative to both Arize AX and Arize Phoenix. It covers tracing, evaluations, prompt management, and experiments in one MIT-licensed product. See Which should you choose.

What is the difference between Arize AX and Arize Phoenix?

Arize AX is the managed platform on Arize's proprietary adb database, launched in 2025; Phoenix is a source-available (Elastic License 2.0) tool you run on SQLite or Postgres. They do not share a storage engine, so Phoenix is not a self-hosted AX. Details in Open source and distribution and History and data plane.

Is Arize Phoenix open source?

Phoenix is source-available under the Elastic License 2.0, which is not an OSI-approved open-source license and restricts offering Phoenix as a hosted service. Langfuse is MIT-licensed.

How does Arize AX pricing compare to Langfuse?

Arize AX Pro is $50/month with 50,000 spans and 10 GB included, then $0.0008 per span and $3 per GB (September 2026). Langfuse bills units from $29/month. At 500k traces / 10 spans / 1 score per trace, that is $621 on Langfuse Pro with 3-year history vs. $4,055 on AX Pro with 30-day retention. Details in the worked example.

Can I self-host Arize AX?

Arize AX self-hosted deployments are listed as Enterprise-only on the Arize pricing page. Phoenix self-hosts freely under the Elastic License. Langfuse self-hosts the MIT product (tracing, evals, prompts, experiments) on every tier.

What does the Dynatrace acquisition mean for Arize users?

Dynatrace announced a definitive agreement to acquire Arize for $915 million on August 13, 2026, expected to close later in the quarter or early in Dynatrace's third fiscal quarter, subject to regulatory review. The announcement does not name Phoenix or state a continuity commitment for a standalone open-source product. Arize AX and Phoenix remain available today.

Did joining ClickHouse change the Langfuse product?

No. Langfuse already ran on ClickHouse. The announcement states the MIT license, self-hosting, Cloud endpoints, and roadmap stay the same, with more capacity to ship. Cloud Core and Pro remain self-serve; OSS self-host has no sales motion. Details in History and data plane and clarifications.

How do I migrate from Arize to Langfuse?

Keep your OpenInference instrumentation and point it at Langfuse's OTLP endpoint, then recreate datasets, prompts, and evaluators. See Switching from Arize.

This comparison is out of date? Please raise a pull request with up-to-date information.