Phoenix vs Langfuse: two OSS-first stacks, two different bets

Phoenix is an Arize project, Apache 2.0, and built on OpenTelemetry from day one. If your platform or SRE team already standardises on OTel as the wire format for traces, Phoenix slots in without teaching anyone a new schema. The instrumentation produces raw OTel spans, which means you can fan them out to any OTel-compatible backend — Tempo, Jaeger, Honeycomb, Datadog APM — with no adapter. Phoenix is strong on evals, especially for teams that already have an eval philosophy and want to drive it from raw spans. The dashboard is lighter than Langfuse, and the product surface is closer to a research tool than a finished product.

Langfuse is MIT-licensed, ships a Docker compose and a Helm chart, and treats its own typed observation model as the source of truth. The dashboard is the best in the OSS category. New SDK features ship every few weeks. Dataset management, prompt versioning, and A/B testing are first-class. OpenTelemetry export is supported, but it is a secondary path: you wire it up if you want it, not the default. The bet is that most teams want a polished product first and an OTel escape hatch second.

Both platforms handle nested traces, tool calls, token usage, and intermediate reasoning. Phoenix is more "open" — the underlying data is raw OTel spans, which is great for queryability and export. Langfuse is more "curated" — the observation objects are typed and the schema is opinionated, which is great for product affordances like prompt versioning and dataset linking. If you want raw access to spans, Phoenix wins. If you want a UI that knows what those spans mean, Langfuse wins.

Both have LLM-as-judge and human review. Phoenix leans academic — the evals draw on Arize research roots, and the evaluator primitives are designed to compose. Langfuse leans product — the dataset UI, prompt versioning, and A/B surface are polished, and the eval runner is wired into the trace UI by default. Neither is objectively better. The right choice is the one your team will actually open every day.

Both ship working Docker compose setups. Langfuse depends on Postgres and benefits from a managed instance in production. Phoenix is lighter — it can run against SQLite for small workloads and against Postgres for production.

If the buyer is your platform or SRE team and they already live in OTel, pick Phoenix. If the buyer is your AI or product team and they ship prompts and datasets weekly, pick Langfuse. If you genuinely cannot decide, run a two-week evaluation of each on a real production trace and pick the one your team actually opens every day. That is the only signal that survives contact with reality.