Arize Phoenix Review (2026)

What Arize Phoenix actually is

Phoenix is Arize’s open-source AI observability and evaluation platform. It does tracing, evals, dataset management, experiments and prompt optimization through a playground, and it is built to run locally - the pitch is a working trace UI on your laptop in under a minute. It is widely called the default open-source choice for LLM observability, and it has earned that with roughly 10.7k GitHub stars and millions of monthly downloads.

Two things to get straight up front. First, Phoenix is the free OSS product; Arize AX is a separate commercial managed platform from the same company. When people say “Arize,” they might mean either. Second, and more important, is the license - because it is not quite what the marketing says.

Arize was founded in 2020 in Berkeley by Aparna Dhinakaran and Jason Lopatecki, and raised a $70M Series C in February 2025 led by Adams Street Partners, with M12, Datadog and PagerDuty among the investors. So the company behind it is well funded and not going anywhere - which matters when you are choosing an observability stack for the long haul.

The gotcha: source-available, not open source

This is the thing to understand before anything else. Arize markets Phoenix as “fully open source, no feature gates.” The main server repo is actually Elastic License 2.0 - source-available, not OSI-approved open source.

The practical difference: ELv2 forbids offering Phoenix “as a hosted or managed service” to third parties. For most teams running Phoenix internally, that restriction never comes up, and the product does behave like open source - you get the source, you can self-host, you can modify it. But it is not the same freedom as MIT or Apache-2.0, and if your business model involves reselling Phoenix as a service, the license rules it out.

There is a real nuance worth stating precisely. The client and eval subpackages - phoenix-evals, phoenix-client, phoenix-otel - are Apache-2.0. The full platform server is ELv2. So the eval libraries you might pull into your own code are genuinely open source; the server you self-host is not. “Fully open source, no feature gates” is marketing that overstates it. The features are not gated - the license is.

What Phoenix does well: OTel and RAG eval

Two capabilities stand out.

It is built on OpenTelemetry and OpenInference. Phoenix is genuinely OTel-native, and Arize maintains OpenInference, its own semantic-convention project for LLM traces. This is the core architectural selling point - it makes Phoenix vendor and framework agnostic, so it works well beyond LangChain, unlike observability tools coupled to one ecosystem. If you are standardizing on OpenTelemetry, this is one of the cleanest fits in the category.

The eval library is deep, especially for RAG. Phoenix ships 50+ pre-built eval metrics and strong LLM-as-judge support, and reviewers single out its RAG evaluation as the best available - you get serious retrieval and answer scoring without writing your own judge prompts. Combined with datasets, experiments and the prompt playground, it is a complete eval workflow, not just a tracing viewer.

The main recurring complaint is ingest performance - there are GitHub reports of a noticeable delay before traces appear in the dashboard, tied to the bulk-inserter path. Worth knowing if real-time visibility matters to you.

Pricing: free OSS, opaque cloud

Tier	Price	What you get
Phoenix (OSS)	$0	Self-host, unlimited spans, all features, ELv2 server
Arize AX	Contact sales	Managed cloud, longer retention, collaboration, support
Arize AX Enterprise	Contact sales	Custom retention, SSO, enterprise support

Phoenix OSS is free with no usage caps and no feature gates on the features themselves. That part is clean.

Arize AX is where it gets murky. The arize.com pricing page was unreadable at verification - it returned an HTTP 403 - so we have no vendor-confirmed AX figures. Third-party aggregators report an AX Pro tier around $50/mo and a free tier metered in spans per month, but those numbers are unconfirmed and we are not going to publish them as fact. Treat AX pricing as contact-sales until Arize publishes it. AX bills in spans per month, not seats or traces.

Phoenix versus Langfuse

The honest comparison is about the license and the eval depth. Langfuse is MIT - genuinely OSI open source - and self-hosting is free with only three features gated. Phoenix’s server is ELv2, source-available, so the freedom is narrower even though the features are not gated. If OSI open source is a hard requirement, Langfuse wins on license alone.

Where Phoenix pulls ahead is eval, especially RAG, and the sub-minute local start. Where Langfuse pulls ahead is the license clarity and a slightly more mature managed-pricing story. Both are OTel-friendly. For a lot of teams this comes down to whether you value Phoenix’s eval depth or Langfuse’s cleaner license more.

Should you use it?

Use Phoenix if you want fast, framework-agnostic OSS tracing, you need strong eval - RAG eval especially - and you are running it for your own use rather than reselling it. The under-a-minute local start makes it the easiest tool in the category to just try.

Think twice if you need OSI-approved open source specifically, or your business model involves offering Phoenix as a hosted service - the Elastic License 2.0 rules that out. And if you are pricing Arize AX, budget a sales call, because the public numbers are not there.

Bottom line: it is the default open-source pick because it is fast, capable and OTel-native, and for internal use the license restriction rarely bites. Just go in knowing “fully open source” is an overstatement - the server is source-available, and that distinction matters for exactly one group of buyers: the ones who wanted to resell it.

---

Pricing and features verified against phoenix.arize.com and the Phoenix repository on 23 July 2026. Arize AX pricing could not be confirmed - the vendor pricing page returned an HTTP 403, and third-party figures are labelled unconfirmed. This category ships breaking changes monthly - we re-verify every 30 days.