Arize AX logo

Arize AX Review (2026)

The commercial cloud product from the makers of Phoenix. The upgrade from free Phoenix is not a tier change - it is a repricing onto per-span billing, and online monitoring is the feature that forces it.

Researched

Rating

4.0

Starting Price

$50/mo

Free Plan

Yes

SDKs & Frameworks

7

Deployment

3

Best For

Teams that outgrew self-hosted Phoenix and want the managed version without operating it, mixed ML and GenAI shops that need computer vision and model observability in the same platform, and regulated teams needing HIPAA.

Last Updated:

10 Things You Should Know About Arize AX

  1. 1 AX Free is 25,000 spans per month with 1 GB storage and 15-day retention
  2. 2 AX Pro is $50/month for 50,000 spans, 10 GB storage and 30-day retention
  3. 3 AX Pro is not charged per seat - cost scales with span volume and data ingestion
  4. 4 The open-source Phoenix build does not include online monitoring in its observability stack
  5. 5 Online evaluations, the Alyx Copilot and enterprise integrations are paid-tier only
  6. 6 One report cites a median Enterprise buyer paying around $60,000 per year
  7. 7 Phoenix's server is Elastic License 2.0, which forbids offering it as a hosted or managed service

Pros & Cons

Pros

  • Not charged per seat - a large team costs the same as a small one at equal volume, which is unusual and favourable
  • Online monitoring and online evaluations against live production traffic, which the free Phoenix build does not provide
  • Inherits genuinely mature ML and computer vision observability from Arize's pre-LLM platform, so mixed ML and GenAI shops get one tool
  • AX Free at 25,000 spans a month is a real evaluation tier, not a demo
  • Phoenix gives you a free, no-lock-in on-ramp before you commit to the cloud

Cons

  • Bills per span, which skews badly against agent workloads where one request can produce dozens of spans
  • Moving from Phoenix to AX is a repricing event rather than a tier upgrade, and reportedly requires schema adjustments and re-instrumentation
  • The gap between AX Pro at $50/mo and Enterprise is large with nothing published in between, and one report puts the median Enterprise buyer near $60,000/year
  • Online monitoring being absent from the open-source build is the single feature most likely to force an unplanned upgrade
  • Phoenix's Elastic License 2.0 restricts offering it as a managed service, which constrains the free escape hatch for platform companies

Features

Managed cloud version of Phoenix with no infrastructure to operate
Online monitoring and online evaluations against production traffic
Alyx Copilot for guided analysis
ML and computer vision observability alongside LLM, inherited from Arize's older platform
HIPAA compliance and formal security review on Enterprise
Usage-based pricing that does not meter seats

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

TierPriceSpans/moStorageRetention
AX Free$025,0001 GB15 days
AX Pro$50/mo50,00010 GB30 days
AX EnterpriseCustomCustomCustomCustom

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.

Pricing Plans

AX Free

$0

  • 25,000 spans per month
  • 1 GB storage
  • 15-day retention
Most Popular

AX Pro

$50/mo

  • 50,000 spans per month included
  • 10 GB storage
  • 30-day retention
  • Not charged per seat - scales with span volume
  • Usage-based overages above the included spans

AX Enterprise

Custom

  • HIPAA compliance and security review
  • Alyx Copilot and enterprise integrations
  • ML and computer vision support
  • Customer success team
  • One report cites a median buyer at around $60,000/year

SDKs & Frameworks

Python SDK TypeScript SDK OpenTelemetry / OpenInference LangChain LlamaIndex CrewAI Any OTel-compatible source

Deployment

Managed cloud (AX) Phoenix self-hosted as the free counterpart Enterprise compliance tiers

Eval Methods

Full evaluation suite Online evaluations (AX only) LLM-as-a-judge Alyx Copilot (AX only) Human annotation

Billing Unit

Spans per month plus data ingestion

Our Verdict

Arize AX is the managed, production-grade half of a genuinely good two-product strategy, and the honest way to evaluate it is to understand what forces the move from free Phoenix. It is not scale, and it is not support. It is online monitoring. The open-source Phoenix build does not include online monitoring or online evaluations as part of its observability stack, so the moment you need to evaluate live production traffic rather than debug locally, you are buying AX. That is a defensible commercial line, and Arize is more upfront about it than most, but it means Phoenix is best understood as an excellent development and debugging tool rather than a complete production platform. AX itself is strong - genuinely mature ML and computer vision observability inherited from Arize's older business, online evals, and pricing that scales on spans rather than seats, which is unusually favourable for large teams. The catch is that per-span billing punishes agent workloads, and the jump from $50 a month to a reported median Enterprise spend near $60,000 a year is a cliff with nothing published in between.

Similar Tools

Frequently Asked Questions

What is the actual difference between Phoenix and AX?

Arize is the company; Phoenix is its open-source LLM observability tool; AX is the commercial cloud product. Phoenix is self-hosted, free and has no usage limits, but no managed features. AX is a cloud product with usage-based pricing, the full evaluation suite and compliance options at Enterprise. The distinction that actually matters commercially is that Phoenix is designed primarily for local development and debugging, while AX targets production-scale deployment. Specifically, online monitoring, online evaluations, the Alyx Copilot and enterprise integrations are AX-only.

What forces the upgrade from Phoenix to AX?

Online monitoring, more than scale or support. The open-source Phoenix build does not offer online monitoring as part of its observability stack, so teams needing production monitoring must move to the paid SaaS version. This is the single most important thing to understand before adopting Phoenix, because it is easy to assume the free product is a complete platform you can grow into. It is an excellent local development and debugging tool. Evaluating live production traffic is a paid capability. Plan for that rather than discovering it after you have standardised.

Is moving from Phoenix to AX easy?

Our sources conflict on this and the disagreement is worth knowing. A competitor analysis characterises it as a repricing event rather than an upgrade, reporting that migrating local Python tracing structures into production cloud infrastructure requires schema adjustments and re-instrumentation. Arize's own materials say the opposite on the technical point - Phoenix and AX share the same OpenInference trace schema, so teams can start locally and migrate without changing instrumentation or data models. On checking further we think Arize has the better of the technical argument, with one caveat - the SDKs differ, with Phoenix using phoenix.otel and an optional API key while AX uses arize.otel with space and API keys, so there is real but modest work. The commercial half of the competitor's framing still stands. You are moving onto per-span billing, and that is a pricing decision rather than a plan setting.

How does per-span billing affect agent workloads?

Badly, and this is the main cost risk. AX bills on spans per month, and agentic requests generate far more spans than simple completions - tool calls, retrieval steps and reasoning chains each produce their own. AX Pro includes 50,000 spans, which is a substantial number of simple completions and a much smaller number of agent sessions. The same criticism applies to Datadog and for the same structural reason. If you run agents, count spans per representative request before you model anything, because a per-request estimate will understate your bill by an order of magnitude.

What happens between $50/month and Enterprise?

Nothing published, and that gap is worth planning around. AX Pro is $50/month, Enterprise is custom, and one report cites a median Enterprise buyer paying roughly $60,000 per year. There is no published tier in between. Usage-based overages above AX Pro's included spans soften the cliff somewhat, but a team growing steadily will at some point have a conversation with sales rather than clicking an upgrade button. Treat the $50 figure as an entry price rather than an indication of what production costs.

Is AX worth it over Langfuse or Logfire?

It depends on whether you need the ML and computer vision side. Arize's pre-LLM business was model observability, and AX inherits genuinely mature capability there. If you run classical ML or CV alongside GenAI, having all of it in one platform is a real advantage that Langfuse and Logfire cannot match. If you are purely an LLM application team, Langfuse self-hosts free with a permissive MIT license, and Logfire is substantially cheaper per unit with a more generous free tier. The per-seat-free pricing model is a genuine point in AX's favour for large teams, though.