What it is
Arize sells two products that share instrumentation. Phoenix is a self-hostable tracing and evaluation app for LLM applications and agents: traces, LLM-as-a-judge and code evaluators, datasets, experiments and a prompt playground. Arize AX is the managed platform: the same tracing and evals at enterprise volume, plus Alyx (an agent that works across your traces and evals), Signal (long-running issue detection), monitors and a self-hosted option. Both ingest OpenTelemetry through Arize’s OpenInference conventions, so an app instrumented for Phoenix can send to AX later without rewriting the instrumentation.
Arize started in classical ML monitoring (drift, feature skew, model performance), and that history is still the reason many teams already have a contract.
Dynatrace now owns it. Dynatrace announced the deal on 2026-08-13 at $915M (about $815M cash plus replacement equity) and completed it on 2026-10-01; its 10-Q records $815.9M cash paid at closing. Co-founder Jason Lopatecki still leads the Arize team, reporting to Dynatrace CEO Rick McConnell. The press release frames the deal as integration, connecting eval results with application performance and infrastructure data; neither it nor Arize’s own post commits to a standalone roadmap, Phoenix’s licence or post-close pricing.
The ops question it answers: you run agents in production (a lead-research agent in RevOps, a contract-intake assistant in legal, a support triage bot) and need to see every tool call, grade the outputs, and find out whether last week’s prompt change made them worse.
What you are actually buying
Phoenix. Released under the Elastic License 2.0. Self-hosting on your own infrastructure is free with no feature gates, according to Arize’s licence page. ELv2 is source-available, not OSI open source: you can run it internally, but you cannot resell it as a hosted service. That matters to a platform vendor, not to an ops team. Phoenix 19.0 and later ships a remote MCP server at /mcp (beta) with OAuth login, read-only SQL and GraphQL queries, and agent code run in a sandbox capped at 30 seconds and 50 operations per call.
Arize AX. Tracing, evals, experiments, human annotation and labelling queues, with unlimited users on every plan. Alyx is the in-product agent: describe the problem and it does root-cause analysis on traces, builds datasets from failures, drafts evaluators and runs prompt experiments, asking before it changes anything unless Auto Accept is on. The hosted AX MCP server is in alpha and exposes create, update and delete on projects, traces, datasets, evaluators, monitors and prompts.
The classical-ML side. Drift and performance monitoring for tabular and recommendation models in the same account. None of Langfuse, LangSmith or Braintrust does this.
Pricing
Published at arize.com/pricing and the AX “Plans and Usage” docs, checked 2026-11-18.
- Phoenix (self-hosted) — $0. No usage limits; you pay for the infrastructure.
- AX Free — $0. 25k spans a month, 1 GB ingestion, 15-day retention, one space, Signal capped at 10 issues a month.
- AX Pro — $50/month. 50k spans, 10 GB ingestion, 30-day retention, one organisation with two spaces, email support, SOC 2 Type II and GDPR.
- AX Enterprise — custom. Custom volume and retention, SaaS or self-hosted, SAML SSO, audit logs, HIPAA, bring-your-own model key for Alyx, custom code evaluators and managed agents, dedicated support.
Data regions on every plan are US, EU and Canada. The current pricing page publishes no overage rate, and pricing trackers report AX Pro’s allowance was cut in Q1 2026 from 100k to 50k spans and from 50 GB to 10 GB.
The real band: AX Pro is a pilot tier. One agent run with 15 spans means 50k spans covers about 3,300 runs a month, so a production agent with real traffic lands on Enterprise or on self-hosted Phoenix. Seats never drive the bill, which is the opposite of LangSmith at $39 per seat per month. For a 12-person team needing SSO, the comparable published numbers are Langfuse Pro plus Teams at $499/month and Braintrust Pro at $249/month flat; on Arize, SSO means an Enterprise quote.
Best for
An ML platform or AI engineering team that already monitors classical models in Arize and now ships agents, and wants one vendor and one instrumentation standard across both. Also a team that wants a free, self-hosted tracing and eval tool it can run on a laptop today (pip install arize-phoenix) and move to a managed platform later. After the deal, also a Dynatrace customer that wants agent traces next to application and infrastructure telemetry.
Not for
A small team that needs SSO and audit logs on a SaaS budget: those features start at an unpublished Enterprise price. Not for a team that requires an OSI-approved licence by policy; Langfuse’s MIT core is the choice there. And not for a team that has not shipped an agent yet: write 20 test cases in a script until there is production traffic to trace.
Versus the alternatives
- Langfuse — pick it when you want MIT-licensed self-hosting, published overage rates and a $29/month Core plan with unlimited users. Langfuse is owned by ClickHouse, so both open-core options now sit inside larger data companies.
- LangSmith — the default for LangChain builders. Pick it when you build on LangChain or LangGraph and want deployment from the same vendor, with fewer than about ten people needing seats.
- Braintrust — the fastest-growing entrant ($800M valuation in February 2026). Pick it when evals are the job, reviewers outnumber engineers, and you want a published flat price.
- Datadog LLM Observability — pick it when the company already pays for Datadog. Arize under Dynatrace is now the same bet on the other APM vendor.
- For voice agents, Coval simulates calls and can import Phoenix traces.
Does the Dynatrace deal change the choice? For Phoenix, not yet: the licence page is unchanged and self-hosting stays free. For AX, the signal to watch is packaging. Dynatrace sells on consumption contracts, and the press release talks about “eliminating fragmentation”, which usually ends in a bundle. Adopt Phoenix freely; sign AX contracts with the protections below.
Watch-outs
- The acquirer has not committed to anything public about AX pricing or roadmap. Guard: cap the renewal increase in the order form, add a termination-for-convenience or data-export clause triggered by a change of product or SKU, and keep instrumentation on OpenInference/OpenTelemetry so traces can be redirected to another backend.
- AX Pro allowances are small and overage is unpublished. At 50k spans, a multi-step agent exhausts the plan in days. Guard: measure spans per run on Phoenix or AX Free for one week, multiply by monthly runs, and get the overage rate in writing before you upgrade; sample low-risk agents.
- The AX MCP server is alpha and writes with your full permissions. Arize’s docs state there is no confirmation step and no dry run. Guard: create a service account scoped to one space for agents, never hand an agent a personal key, and keep production prompts and monitors out of that space.
- ELv2 is not open source in the procurement sense. Guard: have legal confirm internal use is enough, and if you ever plan to offer tracing to your own customers, choose an MIT or Apache-licensed tool.