Comparison · LLM observability

Arize Phoenix alternative: 6 options compared

Phoenix is a local-first tracing and evaluation workbench, not a managed observability product. That makes it excellent for engineers who want control, and awkward for teams that need retention, SSO and a support contract without running infrastructure.

Short answer. Keep Phoenix if you want a free, local-first OpenTelemetry workbench and your team can run it. Choose Langfuse for a self-hostable all-rounder, Braintrust for a polished eval workflow, LangSmith if you are on LangChain, Opik for test suites and agent debugging, or Helicone for request logging and a gateway. Arize AX is the managed upgrade.

What Phoenix is, and what it is not

Phoenix traces LLM and agent runs, stores spans locally, and provides datasets, experiments, evaluations and a prompt playground. It is OpenTelemetry-native, so instrumentation outlives the tool. It is source-available under the Elastic License 2.0, not OSI-approved open source: you may use, copy, distribute and modify it, but not offer it as a hosted service with a substantial set of its features.

That matters for platform teams embedding Phoenix in a developer platform. As a user, the practical questions are simpler: who runs the database, how long do traces live, how do users authenticate, and what happens when volume outgrows one node.

Phoenix is not a model API, a gateway, or a GRC system. It observes the application around the model. Plugsky sits at the model layer; Phoenix or an alternative sits above it.

The six alternatives

OptionLicense and deploymentOpenTelemetryEvalsPrice anchor
LangfuseOpen-source core; cloud or self-hostNative and SDKsDatasets, experiments, LLM-as-judge, annotationFree Hobby; Core $29/mo; Pro $199/mo; Enterprise $2,499/mo
LangSmithCloud; Enterprise self-host/hybridSDK and OTelOffline and online evals, annotation queuesDeveloper $0; Plus $39/seat/mo
BraintrustCloud; Enterprise optionsSDK and OTelStrong experiment and scoring workflowStarter $0; Pro $249/mo
OpikOpen source; cloud or self-hostSDK and OTelTest suites, assertions, 30+ metricsFree OSS and free cloud tier; paid plans at signup
HeliconeOpen source; cloud or self-hostProxy-based logsBasic scores and experimentsHobby free; Pro $79/mo; Team $799/mo
OpenTelemetry + ClickHouse/GrafanaOpen standards; self-host or managedNativeBring your own eval runnerInfrastructure plus engineering time

Feature labels are a planning summary. Licenses, free-tier limits and plan inclusions change; verify each vendor's current documentation and terms.

Self-hosting and the license check

Self-hosting is four requirements: data control, identity, retention and operations. Data control means traces stay in your boundary; identity means SSO, RBAC and audit logs; retention means scheduled deletion; operations means upgrades, backups and incidents.

Phoenix gives you data control and, with work, retention. Identity and operations are yours unless you buy Arize AX Enterprise. Langfuse and Opik include more in their self-hosted editions; Helicone is a logging proxy first.

Elastic License 2.0 is not Apache 2.0. If policy only allows OSI licenses, Phoenix needs an exception or a commercial agreement. Write that down before the pilot, not after.

OpenTelemetry: the migration asset

Instrument once, export anywhere. Use the GenAI semantic conventions for model, operation, token counts and tool calls, and keep vendor-specific attributes separate so you can drop them later.

# Point an OTLP/HTTP exporter at Phoenix
export OTEL_EXPORTER_OTLP_TRACES_ENDPOINT="http://localhost:6006/v1/traces"
export OTEL_SERVICE_NAME="my-agent"
export OTEL_RESOURCE_ATTRIBUTES="deployment.environment=staging"

# The same spans can be sent to Langfuse, Arize AX or a collector
# by changing the endpoint, not the application code.

Phoenix, Langfuse and Arize AX accept OTLP. A ClickHouse and Grafana stack also accepts it, but you build the LLM views, evals and prompt versioning yourself. That suits a platform team with an observability practice; it is a poor first choice for a product team without one.

Evals: what to compare

Look for four capabilities. Datasets need versioning and stable IDs. Experiments need a baseline and diff. Evaluators need code, LLM-as-judge and human annotation, with the judge prompt stored. CI needs a non-interactive command that fails on regression.

Phoenix covers experiments, evaluators and a playground. Langfuse adds prompt management and annotation. Braintrust is the most opinionated experiment workflow. LangSmith is deepest on LangChain. Opik emphasizes test suites and agent debugging. Helicone is lightest on evals.

Keep the judge prompt and dataset version with the application. A score without the evaluator version is not reproducible.

Cost at volume

Phoenix is free; the bill is compute, storage and engineering. Arize AX Free includes 25,000 spans, 1 GB and 15 days. AX Pro is $50/month with 50,000 spans, 10 GB and 30 days; beyond that, overage is a conversation. Langfuse Pro is $199 with 100,000 units; Braintrust Pro $249 with 5 GB and 50,000 scores; Helicone Pro $79 plus usage.

The cheapest option at low volume is usually a free tier. The cheapest at high volume is usually self-hosted storage plus your own compute, if you already have the people. The expensive mistake is choosing a per-span plan for long-retention debugging data and then sampling aggressively to control the bill, which defeats the purpose of tracing.

For a model-layer view of cost, see Plugsky's usage analytics. It reports requests and tokens, not trace-level evals.

Migration

  1. Export spans and datasets. Prefer OTLP or the tool's API over screen-scraping a dashboard.
  2. Map identifiers. Trace ID, span ID, parent ID, session ID and dataset item ID must survive the move.
  3. Re-create evaluators. Port code and judge prompts; re-run a fixed dataset and compare decisions.
  4. Re-instrument only where needed. If the OTel attributes are already correct, change the exporter endpoint.
  5. Run both systems for a sample of traffic and compare completeness and retention.

If LangChain integration is the reason, compare the LangSmith alternatives first. If the eval workflow is too thin, compare the Braintrust alternatives.

When to choose.

  • Phoenix: local-first engineering, OTel, a free workbench, and you can run the infrastructure.
  • Langfuse: self-hostable tracing plus prompt management and evals in one product.
  • LangSmith: you are standardized on LangChain/LangGraph and want the shortest integration path.
  • Braintrust: experiments and scoring are the centre of the release process.
  • Opik: you want test suites, assertions and agent debugging with an open-source core.
  • Helicone: request logging, caching and a gateway matter more than deep evals.
  • OTel plus ClickHouse/Grafana: you already run observability and want one pane for the whole stack.

FAQ

Is Arize Phoenix open source?

Phoenix is source-available under the Elastic License 2.0, not OSI-approved open source. You may use, copy, distribute and modify it, but not offer it as a hosted service with a substantial set of its features. Read the license before commercial embedding.

Is Phoenix free to self-host?

The software is free; you pay compute, storage and engineering. Arize AX Free includes 25,000 spans, 1 GB ingestion and 15-day retention; AX Pro is $50/month with 50,000 spans, 10 GB and 30-day retention.

Does Phoenix require OpenTelemetry?

Phoenix is OTel-native and accepts OTLP traces, but SDKs and integrations also work. Instrument once with GenAI semantic conventions and you can send the same spans to Phoenix, Langfuse, Arize AX or a generic backend.

Which Phoenix alternative is best for self-hosting?

Langfuse and Opik offer self-hosted tracing, evals and prompt management. Phoenix is already self-hosted. Helicone fits when request logging and a gateway matter more than experiments. Check each license and commercial feature split.

Can I migrate Phoenix traces and evals?

Yes, especially with OpenTelemetry. Export spans, datasets and scores, map IDs and attributes, re-create evaluators, and run both systems in parallel. Trace storage moves more easily than the historical meaning of a score.

Sources: Phoenix repository and LICENSE, Arize pricing, Langfuse pricing and Helicone pricing, checked 20 November 2026. This is not legal advice; have counsel review the license before commercial embedding. This page is editorial and does not rank providers by payment.