Skip to main content

Decision brief

Arize Review: Phoenix and Arize AX for Tracing and Evaluating LLM Agents

Start on free, self-hosted Phoenix and keep the option to move the same spans to managed Arize AX, where span volume sets the bill.

Reading time
3 min read
Sources
22 cited
In this piece
Arize+26 Positive

GILD verdict

Choose Arize when your tracing and evals must survive a change of backend, as Bayes Impact's September 2026 pull request did when it moved from Langfuse to Phoenix by repointing a collector.

Self-host Phoenix free first, then move the same spans to Arize AX for managed monitors and compliance. Before production, get AX's overage rate in writing and price it against your span count, or name one owner for Phoenix's database, sign-in, and retention.

Arize

AI Observability by Arize AI

+26 Positive
Product page

Fit and limits

Who it fits

  • Agents are in production or close to it, and the team needs every model call, retrieval, and tool call in one trace when a run goes wrong.
  • Traces must stay in your own network: Phoenix self-hosts with every feature at no license cost and can run air-gapped.
  • The team wants instrumentation it can move: OpenInference spans go to Phoenix, Arize AX, or other backends such as Datadog without rewriting the application.
  • The same team also monitors classification, regression, ranking, or vision models for drift and performance, which Arize logs alongside LLM traces.
  • A flagged failure must be right before a person acts on it: in Parlance Labs' July 2026 test, 91% of the failures Alyx flagged on 100 production traces were valid, and three were wrong.

Where it falls short

  • AX Free allows 25,000 spans a month and Pro allows 50,000, and Arize's pricing page lists no Pro overage rate as of Oct 28, 2026. A production agent's span count sets the bill, so get that rate in writing before you rely on the base fee.
  • A Phoenix instance is one tenant with instance-wide roles, so separate teams need separate instances. Arize's architecture docs plan access scoped to projects, through resource tags, for later in 2026.
  • Your team handles Phoenix at volume. Public accounts report slow attribute search at about 100,000 interactions a day and PostgreSQL deadlocks that lost traces. Arize scales Phoenix by adding instances on one database and sends high-volume ingestion to AX.
  • Phoenix is under the Elastic License 2.0, which bars offering it to third parties as a managed service and is not an OSI-approved open-source license. A bbv developer proposed moving its AI Hub to MIT-licensed Langfuse for that reason in December 2025, and the team merged the switch in February 2026.
  • Phoenix can become one service too many. Elastic dropped it from Kibana's eval framework in August 2026 once datasets lived in Elasticsearch, and Apify dropped its Phoenix eval runner in September 2026 after Langfuse-based gates replaced it.

Implementation reality

  1. Instrument once with OpenInference

    Register framework and provider OpenInference instrumentors with arize-phoenix-otel or arize-otel. Name a project per service. When migrating, send identical spans to Phoenix and AX for one traffic cycle, as Arize recommends.

  2. Mask before export

    Drop inputs and outputs with OPENINFERENCE_HIDE_INPUTS and OPENINFERENCE_HIDE_OUTPUTS. Use finer switches for messages and images, plus embeddings. Redact PII in a span processor with regex or Microsoft Presidio before export.

  3. Decide where Phoenix stores data

    Use PostgreSQL 14 or later for a shared Phoenix instance. Default SQLite data disappears if the container has no volume. Enable OAuth2 or LDAP authentication. From Phoenix 9.0, set retention per project.

  4. Gate changes and watch production

    Save failing traces as datasets. Run code and LLM-judge experiments on each prompt or model change. On AX, set an online-eval sampling rate and monitors that notify Slack, PagerDuty, Opsgenie, or Microsoft Teams.

What to measure in your pilot

  • Count spans per agent run for a week of real traffic, ask Arize for Pro's overage rate in writing, and stay on Pro if that beats an Enterprise quote.
  • Dynatrace says AX and Phoenix stay available and roadmap updates will follow; track Phoenix's monthly release count from October 2026 and plan a move if it falls below half its 2026 average.
  • Load-test Phoenix on PostgreSQL at twice your peak daily spans for a week; if attribute search passes 5 seconds or spans drop, move that workload to AX.

Sources

Official documentation
22
  1. Arize AX pricingOfficial documentationarize.com
  2. Arize AX docs: Pricing and usageOfficial documentationarize.com
  3. Arize AX docs: Release notes (13-month retention on Pro, Jul 10, 2026; Signal general availability, Jul 29, 2026)Official documentationarize.com
  4. Arize AX docs: Self-hostedOfficial documentationarize.com
  5. Arize AX docs: Arize Private ConnectOfficial documentationarize.com
  6. Arize AX docs: ComplianceOfficial documentationarize.com
  7. Arize AX docs: SSO & RBACOfficial documentationarize.com
  8. Arize AX docs: Mask and redact sensitive dataOfficial documentationarize.com
  9. Arize AX docs: Set up tracingOfficial documentationarize.com
  10. Arize AX docs: Run online evals on tracesOfficial documentationarize.com
  11. Arize AX docs: Continuously monitorOfficial documentationarize.com
  12. Arize docs: ML models (predictions, drift, and performance)Official documentationarize.com
  13. Arize AX docs: Migrate from Phoenix to Arize AXOfficial documentationarize.com
  14. Phoenix docs: LicenseOfficial documentationarize.com
  15. Arize Phoenix LICENSE (Elastic License 2.0)Official documentationgithub.com
  16. Phoenix docs: Self-hostingOfficial documentationarize.com
  17. Phoenix docs: ArchitectureOfficial documentationarize.com
  18. Phoenix docs: Access control (RBAC)Official documentationarize.com
  19. Phoenix docs: Data retentionOfficial documentationarize.com
  20. Phoenix docs: PrivacyOfficial documentationarize.com
  21. Dynatrace: Dynatrace completes acquisition of Arize (Oct 1, 2026)Official documentationdynatrace.com
  22. Arize: Compare ArizeOfficial documentationarize.com