Arize AX Review (2026)

Two products, one company, and the line between them

Getting this straight first, because the naming causes real confusion.

Arize is the company. Phoenix is its open-source LLM observability tool, free and self-hosted. Arize AX is the commercial managed cloud product. When people say “Arize,” they could mean either product.

AX is the enterprise SaaS version of Phoenix, adding the Alyx Copilot, ML and computer vision support, HIPAA compliance, formal security review and a customer success team.

The interesting question is not what AX adds. It is what makes the free version insufficient - and the answer is more specific than most comparisons acknowledge.

What actually forces the upgrade

Not scale. Not support. Online monitoring.

The open-source Phoenix build does not offer online monitoring as part of its observability stack. Online evaluations, the Alyx Copilot and enterprise integrations are paid-tier only. Phoenix is designed primarily for local development and debugging; AX targets production-scale deployment.

This matters because of how teams adopt Phoenix. It runs locally in under a minute, it has no usage caps, and it has no feature gates on the things you touch while developing - tracing, datasets, experiments, the playground. It feels like a complete platform. Teams standardise on it.

Then they go to production, want to evaluate live traffic rather than replay local traces, and discover that capability is on the other side of a purchase.

That is a defensible commercial line - Arize has to make money somewhere, and gating production monitoring rather than crippling the developer experience is a more honest choice than most. But plan for it up front. Phoenix is an excellent development and debugging tool. Treating it as a free production observability platform sets you up for an unplanned migration at the worst moment.

Pricing

Tier	Price	Spans/mo	Storage	Retention
AX Free	$0	25,000	1 GB	15 days
AX Pro	$50/mo	50,000	10 GB	30 days
AX Enterprise	Custom	Custom	Custom	Custom

Two genuinely good things here.

AX Pro is not charged per seat. Cost scales with span volume and data ingestion, not headcount. This is unusual and it is favourable - a twenty-person team pays the same as a three-person team at equal volume. Compare LangSmith, where seats are the primary meter. For growing teams this is a meaningful structural advantage.

AX Free at 25,000 spans a month is a real tier, enough to evaluate properly rather than a demo.

And two problems.

Per-span billing punishes agents. AX bills spans, and an agentic request produces many - tool calls, retrieval, reasoning steps. The 50,000 included spans on Pro is a lot of simple completions and rather few agent sessions. This is structurally the same trap as Datadog’s, and the same mitigation applies: count spans for a representative request before modelling anything.

The gap above Pro is a cliff. Nothing is published between $50/month and Enterprise, and one report cites a median Enterprise buyer at around $60,000 per year. Usage-based overages above Pro’s included spans soften the edge, but at some point growth means a sales conversation rather than an upgrade button. Treat $50 as an entry price, not a production cost.

The migration is a product change, not a plan change

Our sources disagree here, so both are worth stating.

A competitor analysis frames graduating from Phoenix to AX as a repricing event rather than a tier upgrade, reporting that migrating local Python tracing structures into production cloud infrastructure requires schema adjustments and re-instrumentation.

Arize’s own materials contradict the technical half. Phoenix and AX share the same OpenInference trace schema, so teams can start locally with full infrastructure control and later migrate without changing instrumentation or data models.

On further checking we think Arize has the better of the technical argument, with one caveat: the SDKs differ - Phoenix uses phoenix.otel with an optional PHOENIX_API_KEY, while AX uses arize.otel with space and API keys. That is real but modest work, not a re-instrumentation project.

The commercial half of the competitor’s framing still stands. You are moving onto per-span billing, which is a pricing decision rather than a plan setting - and that remains the substantive thing to plan for.

That analysis comes from a competitor and the framing is motivated, so discount it accordingly. But the underlying point holds: you are moving between two products that share a lineage, not changing a setting. Budget engineering time.

There is also a constraint on the free escape hatch worth knowing. Phoenix’s server is Elastic License 2.0, which forbids offering it as a hosted or managed service to third parties. For internal self-hosting this is a non-issue. For a platform company or agency wanting to build a product on top of Phoenix rather than pay for AX, it closes that door - and some legal teams reject ELv2 outright since it is not OSI-approved open source.

Where AX is genuinely strong

The differentiator most comparisons undersell: Arize’s pre-LLM business was model observability, and AX inherits mature ML and computer vision capability from it.

If you run classical ML models or CV alongside GenAI, AX covers all of it in one platform. Langfuse cannot do this. Logfire cannot do this. For a data science organisation that has been doing ML for years and is now adding LLM features, that consolidation is worth real money and is the strongest reason to choose AX over a purely LLM-native tool.

Should you use it?

Use AX if you have outgrown self-hosted Phoenix and need online monitoring, you run ML or computer vision alongside GenAI, you need HIPAA, or your team is large enough that per-seat pricing elsewhere hurts.

Don’t use it if you are a pure LLM application team on a budget - Langfuse self-hosts free under MIT and Logfire is cheaper per unit - or you run agent-heavy workloads where per-span billing works against you.

Bottom line: a strong managed platform with an honest free on-ramp, let down slightly by a pricing cliff and a billing unit that suits completions better than agents. The thing to internalise before adopting Phoenix is that online monitoring lives on this side of the paywall. Decide with that in view rather than discovering it in production.

---

Pricing and the Phoenix/AX feature split verified against vendor documentation and third-party analyses on 31 July 2026. The median Enterprise figure is a single third-party report and is flagged as such. This is a researched directory entry - we have not yet instrumented this platform with our reference application.