What it does
Arize AX is Arize's AI engineering platform for observing and improving LLM applications and agents. Teams send traces from instrumented applications, inspect spans, use Signal to surface recurring issues, and apply evaluations, experiments, dashboards, and monitors to production behavior. Its published plans include a hosted free tier and paid options, with self-hosted deployment listed for Enterprise.
Start with a representative trace
Arize AX's documented workflow begins by instrumenting an LLM application or agent with an integration or OpenTelemetry-based instrumentation, then sending traces to a project. A trace records the steps in a request; span details can show inputs, outputs, latency, and token or cost information when those fields are captured. Signal can help surface recurring issues in production traces, while evaluations and experiments let a team check changes against criteria it defines. These are documented capabilities, not results from an instrumented app in this research.
Turn an observed metric into a monitor
For production monitoring, the docs describe selecting a tracing project, choosing a span attribute or custom metric, setting an evaluation window and threshold, and configuring notifications. Examples include latency, token counts, and evaluation labels. Dashboards provide a broader view, but monitors are the notification path when a tracked measure crosses its threshold. The default schedule for a tracing-project monitor is hourly, so check its frequency when the response time matters.
AX's pricing page lists a free SaaS plan with 25,000 spans and 1 GB of ingestion per month, 15-day retention, and no credit card requirement. Pro is listed at $50 per month with larger span and ingestion allowances; Enterprise pricing is custom and may use SaaS or self-hosted deployment. The published limits are a starting point for evaluation, not a complete estimate for a high-volume workload. No account was created and no trace was sent.
Strengths
- Traces and span trees help developers inspect the steps, inputs, outputs, latency, and errors behind AI application requests.
- Custom monitors can evaluate span attributes or custom metrics against thresholds and send notifications.
- Signal, evaluations, and experiments provide separate ways to find recurring issues, score behavior, and check changes.
- The free hosted plan publishes span, ingestion, and retention limits, while Enterprise lists a self-hosted option.
Watch out for
- Tracing-project monitors run hourly by default, so teams should configure a suitable schedule rather than assume immediate evaluation.
- Monitoring depends on application instrumentation and incoming traces; the account, keys, and provider setup were not exercised here.
Best for
- AI engineering teams investigating which model, tool, or application step caused a problematic response.
- Teams alerting on production LLM latency, cost, token use, or evaluation results.
- Organizations evaluating hosted AI observability with a published free tier before considering Enterprise self-hosting.
Sources
- Arize AX - AI Engineering Platform for Self-Improving AI Agents (Oct 26, 2026)
- Arize AX Pricing: AI agent observability and evaluation - Arize AI (Oct 26, 2026)
- Arize AX - Arize AX Docs (Oct 26, 2026)
- Continuously Monitor - Arize AX Docs (Oct 26, 2026)