Arize Phoenix is an open-source platform for AI observability and evaluation, built by Arize. It traces each step an agent takes so developers can see where a run went wrong, and it provides evaluations, experiments and annotation tools for measuring quality. Phoenix runs locally with a single command or can be self-hosted, and a managed route exists through Arize AX. The AX Free plan includes 25,000 spans per month, AX Pro costs $50 per month, and Enterprise pricing is custom.
Phoenix serves AI engineers, data scientists and ML teams who build LLM applications and want visibility without sending data to a closed service. Individual developers can start locally in seconds, while larger organizations can graduate to Arize AX for SaaS or self-hosted deployments with higher limits. Teams that rely on frameworks such as LlamaIndex benefit from one-click integration, and those standardizing on OpenTelemetry can use its OpenInference instrumentation. Reviewers who are not engineers can add annotations and labels to traces to build quality datasets.
A developer investigating a hallucination can open a trace to inspect retrieval results and tool calls that led to a bad answer. A team tuning a prompt can run experiments against a dataset and compare results with evidence instead of guesswork. Evaluation workflows use one model to score another for relevance, toxicity and response quality. Support-focused teams analyze unstructured text to find root causes of unexpected user inputs and gaps in a knowledge base. A coding agent can be used to add tracing, and the in-product assistant helps users talk through their traces to investigate issues and run experiments.
