Arize AI review

Arize AX is a managed SaaS platform for observing, evaluating and improving AI applications, including agents. Instrument an application, send traces, then use the hosted product to inspect agent activity, run evaluations and investigate production issues. Traces include token, latency and cost data; evaluations can run at span, trace and session levels. Signal surfaces issues and proposes fixes, while Alyx is an AI engineering agent for running evaluations and debugging.

Phoenix is Arize’s open-source, local-first observability and evaluation option. The vendor says it can run locally or be self-hosted, letting teams trace LLM calls, evaluate behavior and investigate failures. Both products use OpenInference and OpenTelemetry conventions. AX adds managed infrastructure, online evaluations and workflows for production improvement; its Enterprise plan lists SaaS or self-hosted deployment. In a stack, Arize sits alongside the application and its model or agent framework, collecting traces for development and production analysis.

The documented integration path is to connect an agent or application and send its first trace. Arize names OpenInference, built on OpenTelemetry, as its semantic convention and tracing foundation; the stated aim is instrumentation without a proprietary trace format. The homepage offers a terminal setup through npx and links to a setup guide, docs and cookbooks. It does not name programming languages or describe an API style, so confirm SDK and API support in Arize’s docs before choosing an instrumentation path.

The vendor lists integrations with more than 40 models, frameworks and AI tools, including OpenAI, Anthropic, Google, Amazon Bedrock, LangGraph, LangChain, LlamaIndex, CrewAI, OpenAI Agents SDK and DSPy. It also names Cursor, Claude Code and OpenCode for agent-native workflows. Open standards support work across Google Cloud, AWS, Azure and self-hosted environments. ADB stores traces in open formats and connects through DataFabric to BigQuery, Databricks and Snowflake. Check the docs for the exact integration steps and compatibility for your stack.

AX Free costs $0 per month and includes 10 issues, 25,000 spans, 1 GB of ingestion and 15-day retention per month. It includes unlimited users and evaluations, but its issue, span, ingestion and retention ceilings make it a trial-sized production allowance rather than an unlimited free service. Phoenix is the open-source local-first option; the published pricing does not list a Phoenix fee.

AX Pro costs $50 per month and raises the listed allowances to 25 issues, 50,000 spans, 10 GB ingestion and 30-day retention, with unlimited users and evaluations. The plan price is monthly; annual pricing is not published. AX Enterprise has custom pricing, unlimited Signal issues and custom spans, ingestion and retention, plus SaaS or self-hosted deployment. The pricing page does not list overage rates, add-ons or a free trial. Before production, ask how additional usage is handled, whether limits can be raised mid-cycle and what Enterprise deployment costs. User counts and evaluation counts are unlimited on the listed AX plans, so the published thresholds center on usage and retention.

Phoenix is described by Arize as open source and can run locally or be self-hosted. The vendor does not name Phoenix’s licence on the vendor's site, so check the repository licence and obligations before distributing or modifying it. AX Enterprise lists SaaS or self-hosted deployment; AX Free and Pro are listed as monthly plans without a self-hosted option.

Arize says OpenInference uses OpenTelemetry and avoids a proprietary trace format. Its ADB datastore stores traces in open formats and connects through DataFabric to BigQuery, Databricks and Snowflake. Those choices can make instrumentation and downstream access more portable, but they do not by themselves establish complete export coverage or a migration path for evaluations, dashboards and workflow configuration. Confirm which data and configuration can be exported, in what formats, and whether export carries a fee. Phoenix offers a route to run without AX’s managed service, but expect to operate the infrastructure yourself.

AX Free and Pro impose explicit monthly ceilings for issues, spans and ingestion, with retention of 15 and 30 days respectively. Enterprise allows custom spans, ingestion and retention. The vendor does not publish rate limits or SLA terms on its homepage, so ask about ingestion behavior at peak load, alert delivery, incident response and service commitments before making AX a production dependency.

Arize lists SOC 2 Type II, ISO 27001, PCI DSS, HIPAA and GDPR among its security and compliance credentials. Verify the specific deployment’s scope, data handling terms and requirements with the Trust Center. One Reddit thread argues that tracing prompts and completions may not reveal infrastructure causes such as Kubernetes OOMs; a commenter points out that infrastructure monitoring can cover those failures. That is a useful boundary to plan for: agent traces explain application behavior, while cluster, network and database monitoring may still be needed to diagnose the full request path.