What it is
Arize started as a leading ML model monitoring platform — drift detection, model performance, tabular data — and extended into LLM observability as the generative wave hit. That extension included the open-source Phoenix project, which has become a widely-used standalone trace viewer for LLM apps.
Free tier (Phoenix): 25K trace spans/month, single user. Cloud plans require sales contact for production capacity.
Where it shines
- ML + LLM in one platform. If you're operating both in production, the unified view is real and rare.
- Production maturity. Arize's ML monitoring has been in front of regulators and procurement teams for years. That credibility carries.
- Phoenix. The OSS trace viewer is genuinely good and a credible way to "try before you buy."
Where it falls short
- ML-first architecture. The data model and workflow assumptions show their origins. LLM-native features (prompt management, dataset-driven evaluation, CI gates) feel grafted on.
- No deploy blocking. Evaluation results are surfaced; they don't gate releases.
- Workflow fragmentation. Datasets, traces, and evaluations live in separate parts of the product.
Bottom line
If your AI footprint is both classical ML and LLMs, Arize is the most credible unified play in the market. If your work is LLM-only, the eval-first specialists (Braintrust, Langfuse) deliver more LLM-specific value with less workflow friction.