Pick your LLM-observability tool on two questions, not on a feature grid — because on features these three converge. All of them do distributed tracing, LLM-as-judge and code evals, versioned prompt management, and datasets. The differences that actually change your life are: what can you legally self-host for free, and is your telemetry portable or trapped in one vendor's model.
Here's the short answer before the detail. Want the full product for free on your own box? Langfuse. Standardizing on OpenTelemetry so you can also ship traces to Honeycomb or Datadog? Arize Phoenix. Live or die by whether a prompt change regressed quality? Braintrust.
Langfuse — the default self-host#
Langfuse (YC W23, acquired by ClickHouse in January 2026) is the one to reach for if "own my data, pay nothing" is the goal. The core is MIT-licensed, it has the largest community of the three (~32k stars), and — critically — the self-hosted build is not a crippled demo. Tracing, evals, prompt management, datasets, and the LLM playground are all in the open core. Only thin enterprise-compliance features (SCIM, audit logs, fine-grained RBAC) sit behind a commercial license. The acquisition didn't change that; the license and the no-gate self-host held.
If you never want to self-host, the cloud tiers are Hobby (free, 50k units/month), Core ($29/month), and Pro ($199/month), where a "unit" is a trace, observation, or score. SDKs are Python and JS/TS, with first-class integrations for the OpenAI SDK, LangChain, LlamaIndex, LiteLLM, and the Vercel AI SDK. Langfuse also ingests OpenTelemetry, so it plays nicely with the standard even though its own SDK is the fast path.
On features these three converge. They split on license and on whether your traces can leave.
Arize Phoenix — the OpenTelemetry-native pick#
Phoenix, from Arize AI, is the choice when portability is the priority. It's built OTel-native: it emits and reads the OpenTelemetry GenAI semantic conventions, the schema the industry is converging on for LLM and agent telemetry. That's the whole point — instrument once against the standard, and the same spans can flow through an OTel Collector to Phoenix and to Honeycomb's Agent Timeline or Datadog, with no code rewrite. Phoenix also has the broadest SDK coverage here (Python, JS/TS, plus Java and Go) and unusually strong built-in evals for RAG failure modes — faithfulness, relevance, hallucination, toxicity.
The catch is the license. Phoenix's server is Elastic License 2.0 — source-available, not OSI-approved, and it forbids offering Phoenix as a hosted service to third parties. For a founder self-hosting for internal use, that restriction almost never bites: local and your-cloud deployment is free with no feature gates. But if you were planning to resell a hosted Phoenix, read the license first.
If you want the mechanics of emitting standard traces so any of these backends can read them, we wrote the how-to: trace your agent with OpenTelemetry GenAI, then point it anywhere.
Braintrust — when evals are the job#
Braintrust is the odd one out, and deliberately so. The platform is proprietary SaaS (only the client SDKs are Apache-2.0), and self-hosting is Enterprise-only. You choose it not for openness but because evals are its entire reason to exist: scoring is the metered unit, and the experiment, regression-testing, and human-review workflow is the most polished of the three. If your real question every week is "did that prompt edit make the product worse," Braintrust answers it with the least friction. The free Starter tier (1 GB, 10k scores) is enough to prove that out; Pro is $249/month, and costs scale with GB of processed data and thousands of scores.
The decision#
- Self-host Langfuse if you want to own your data, pay nothing, and get the full feature set. This is the right default for most one-person teams.
- Start with Phoenix if you're already emitting OpenTelemetry, or want to keep telemetry vendor-neutral so you can fan it out to Honeycomb/Datadog later.
- Add Braintrust when eval velocity — not tracing — becomes your bottleneck, and a proprietary SaaS is an acceptable trade.
And the incumbent: LangSmith is the frictionless choice only if your stack is already all-in on LangChain. Wanting to own your data or dodge framework lock-in is precisely why people leave it — for one of the three above.
One more thing that ages well: whatever you pick, instrument against the OpenTelemetry GenAI conventions where you can. Tools get acquired, relicensed, and repriced — a ClickHouse acquisition here, an Elastic License there. Standard traces are the one asset that survives a backend swap.
