Arize AX

Arize AX is a managed AI engineering platform that traces, evaluates and improves LLM applications and agents. It captures every model call, retrieval and tool invocation as an OpenTelemetry trace, scores those traces with code, human or LLM-as-a-judge evaluators, and clusters recurring production failures into ranked issues so AI teams can prove a fix before shipping it.

Arize AX is a managed AI engineering platform for tracing, evaluating and improving LLM applications and agents across development and production. It captures OpenTelemetry-based distributed traces of every agent run — model calls, document retrievals, tool invocations, inputs, outputs, latency and token cost — using the OpenInference semantic conventions Arize authored, then attaches evaluators to score those traces at span, trace and session level: deterministic code checks, human annotations, LLM-as-a-judge and agent-as-a-judge. Signal automatically groups recurring production failures into ranked issues rather than leaving engineers to read traces by hand; Experiments run controlled comparisons so a prompt, model or retrieval change can be proven against a dataset before it ships; and Alyx is an embedded AI agent that answers questions about trace data in plain language, runs evaluations and drafts fixes, with agent-native workflows exposed for Cursor, Claude Code and OpenCode. AX sits alongside Phoenix, Arize's open-source, self-hostable tracing and evaluation library, which the company reports at over two million monthly downloads and which carries roughly 11,300 GitHub stars — a deliberate bottom-up motion in which developers adopt the free tool before enterprise procurement begins. The platform supports 40+ models and frameworks including OpenAI, Anthropic, Google, Amazon Bedrock, LangChain, LangGraph, LlamaIndex, CrewAI and DSPy, running on AWS, Google Cloud, Azure or self-hosted. Founded in January 2020 by Jason Lopatecki and Aparna Dhinakaran and based in Berkeley, California, Arize raised a $70M Series C in February 2025 led by Adams Street Partners, taking total funding to $131M. On 13 August 2026 Dynatrace agreed to acquire the company for approximately $915M in cash and stock.

Category

Governance & Security

Pricing

Freemium, Subscription, Usage-based, Contact for pricing

Target Market

CTOs, Data Scientists, Enterprise Developers, ML Engineers, AI Platform Teams

Deployment

Cloud-first, Self-hosted, Open-source, Hybrid, Multi-cloud

Founded

2020

Headquarters

Berkeley, California, United States

Team Size

201-500

Customers

Hundreds of enterprise and government customers, publicly named including Uber, Duolingo, Booking.com, Priceline, TripAdvisor, PepsiCo, Hyatt, Wayfair and Condé Nast

• ✓
  OpenInference agent tracing

  Captures every model call, retrieval and tool invocation of an agent run as one OpenTelemetry trace, so a multi-step failure can be replayed rather than guessed at from logs.

• ✓
  Multi-modal evaluation

  Scores spans, traces and sessions using code evaluators, human annotations, LLM-as-a-judge and agent-as-a-judge, letting teams grade subjective output quality without hand-reviewing every run.

• ✓
  Signal issue clustering

  Automatically groups recurring production failures into ranked issues, which turns a firehose of traces into a short prioritised list an engineer can actually work through.

• ✓
  Experiments

  Runs controlled comparisons of prompt, model or retrieval changes against a fixed dataset, so a change is proven to improve quality before it reaches production users.

• ✓
  Alyx AI engineering agent

  An embedded agent that answers questions about trace data in plain language, runs evaluations and drafts fixes, cutting the expertise needed to interpret observability output.

• ✓
  Phoenix open source

  A self-hostable tracing and evaluation library built on OpenTelemetry with over two million monthly downloads, so teams can start locally with full data control and no vendor lock-in.

• ✓
  Broad framework coverage

  Auto-instruments 40+ models and frameworks including OpenAI, Anthropic, Bedrock, LangGraph, LlamaIndex, CrewAI and DSPy, so instrumentation survives a change of model provider or agent framework.

• •
  Root-causing an agent that fails intermittently

  Replay the full trajectory of a failed run to find whether the model, the retrieval step or a tool call broke the chain.

• •
  Catching hallucination and quality drift in production

  Run continuous LLM-as-a-judge evaluators against live traffic and alert when groundedness or relevance scores degrade against baseline.

• •
  Validating a prompt or model migration

  Compare the candidate configuration against the incumbent on a golden dataset and ship only if measured quality holds or improves.

• •
  Controlling LLM spend

  Attribute token cost and latency down to individual spans to find the retrieval step or verbose prompt driving most of the bill.

• •
  Evidencing AI quality for governance and audit

  Retain scored traces of production decisions so risk and compliance teams can review what the system did and why.

Unusually for this category, seats are not the meter — all three tiers include unlimited users, evaluations, experiments and human annotations, and consumption is billed on span volume, storage, retention window and Signal issues. AX Free covers 25k spans and 1GB a month at 15-day retention; AX Pro is $50/mo for 50k spans, 10GB and 30-day retention. That span ceiling is the number to model against, because a verbose agent emits many spans per user request and PeerSpot reviewers report costs climbing with heavy tracing and large embeddings. Everything an enterprise buyer typically requires sits behind unquoted Enterprise pricing: unlimited spans, custom retention, self-hosted deployment, multi-org RBAC, SSO, audit logs, HIPAA and support SLAs. The open-source Phoenix library remains a genuinely free self-hosted path for teams willing to run and store it themselves.

This page was written from 8 sources, 5 on domains other than arize.com.

1. 1.arize.com — arize.comvendor
2. 2.arize.com — pricingvendor
3. 3.arize.com — arize ai raises 70m series c to build the gold standard for vendor
4. 4.peerspot.com — arize ai reviews
5. 5.forbes.com — dynatrace pays 915 million to move ai evaluation upstream
6. 6.dynatrace.com — dynatrace to acquire arize
7. 7.constellationr.com — dynatrace acquires arize accelerates ai observability effort
8. 8.github.com — phoenix