Arize AI

• Dynatrace's agreement to buy Arize is subject to regulatory review and has not closed, so the vendor a buyer signs with today may be a Dynatrace business line by the time the contract renews.
• Datadog ships its own agent observability product and was named a strategic backer of Arize's 2025 Series C, so a monitoring platform an enterprise already buys could absorb the same tracing and evaluation.
• The free Phoenix library covers the tracing and evaluation the paid platform sells, so teams that adopt Phoenix may never move onto a paid plan.
• Arize traces calls to model providers including OpenAI, Anthropic and Google, any of which could evaluate the agents built on its own platform and remove the third-party budget line Arize sells into.
• Arize instrumentation follows OpenTelemetry, an open standard, so a customer's tracing code can feed a replacement without a rewrite.
• Runtime guardrails are one part of an engineering platform, so a buyer shopping specifically for AI security may pick a vendor that sells only that.

Arize sells engineers a way to see and score what their AI systems actually did. The homepage names chatbots, retrieval systems, copilots and agents as what teams build with it, and says teams use Arize to understand how those systems behave and to improve them. The platform records each step a model or agent takes and then scores the result. Runtime guardrails extend the same machinery to run-time, correcting or blocking a message that fails a guard.

Independent analysts now describe the same problem in the same terms. Gartner opened a Market Guide for AI Evaluation and Observability Platforms in February 2026 and told software engineering leaders to invest in the category, on the grounds that nondeterministic generative AI is hard to measure and hard to trust. Its March 2026 release puts the tooling in the hands of the teams that build and operate AI systems and, increasingly, the operations engineers who keep them running.

Security is the narrow part of the story. Arize AX documents two guard types, one that flags inputs resembling supplied examples of bad messages and one that judges retrieval-augmented answers, and the default guard ships with a public jailbreak dataset behind it. Arize presents the platform as an engineering tool for teams improving AI quality, so the guardrail is one capability inside it rather than the whole product. [s1, s2, s3, s14, s15]

The platform spans tracing, evaluation, experimentation and runtime guardrails rather than a single control. Arize records the steps an agent takes, runs evaluations over those recordings, and gives a team a playground to optimize prompts, compare models and replay traced calls. It is built on OpenInference and OpenTelemetry standards and runs across Google Cloud, AWS, Azure and self-hosted environments.

Phoenix is the part a buyer can inspect directly. The open-source repository documents tracing, evaluation, datasets, experiments, playground and prompt management, is vendor and language agnostic with support for frameworks including LangGraph, CrewAI and LlamaIndex, and carries 11,174 stars and 1,074 forks. That open code is what a reader can check instead of a third-party benchmark.

The guardrails are defined but bounded. The documentation describes guards that correct undesirable outputs at run-time and trigger default responses, retries or blocking, and it states that a customer can instantiate a guard with off-the-shelf prompts and datasets or supply its own, with any model provider behind it. [s1, s2, s3, s5, s7, s8]

Arize competes against independent evaluation vendors, monitoring incumbents and its own free tool. TechCrunch describes an increasingly crowded observability and evaluation space and names Galileo and Patronus AI as very similar offerings. Datadog sells its own agent observability product and was named a strategic backer of Arize's 2025 Series C.

Its visible differentiator is open-source reach with no lock to one model vendor. Phoenix is vendor and language agnostic, the tracing follows OpenTelemetry, and Constellation Research describes a strong open-source community working across all major AI frameworks. That openness is what Arize offers a buyer choosing it over a bundled monitoring suite.

Who owns the workload is the question the Dynatrace agreement puts in play, and the answer would change on closing. Dynatrace signed a definitive agreement to buy Arize for 915 million dollars, so the independent position Arize sold against monitoring incumbents becomes, on closing, part of one. Constellation Research reads the deal as observability vendors moving to own the record of what an AI decided. [s7, s9, s12, s13, s19]

Open-source adoption is the top of Arize's funnel, and what happens below it is not disclosed. Arize's own pricing page calls Phoenix the open-source, local-first platform and tells a reader to move to Arize AX only when they need it, which is a funnel by design rather than a measured conversion. Co-founder Aparna Dhinakaran told TechCrunch that the open-source product may be Arize's biggest competitor and that the company planned to spend more on it.

Named accounts sit alongside that free base. TechCrunch reports that Arize works with enterprises including Uber, Klaviyo and Tripadvisor, and the customers page carries an LG U+ case study whose named team lead describes evaluation-driven development for 30 million subscribers. None of those sources records whether an account arrived through Phoenix or bought directly, so the named logos do not size the conversion.

The federal record is the piece a reader can check without the vendor. USAspending records two Department of Defense purchase orders to Arize AI Inc. The Air Force Research Laboratory placed one for 1,196,635 dollars for language-model evaluation and development tooling, running to July 2026, and the Navy's NIWC Pacific placed a 451,939 dollar order for AI machine learning observability in 2024 that the Department of the Air Force funded. [s1, s5, s7, s9, s16, s17, s21]

One founder had led a machine-learning startup that Arize absorbed at launch, and the other came from a company Adobe bought. Aparna Dhinakaran built machine learning infrastructure at Uber and was chief executive of Monitor ML, a Y Combinator-backed startup Arize acquired at launch in 2020, so a founder with a prior build in the same problem space was inside the company from day one. Jason Lopatecki came from TubeMogul, the video advertising company Adobe bought for over 500 million dollars in 2016.

The bench around them is built out for an enterprise motion. The about page lists a chief technology officer, a chief information security officer, a compliance officer, a head of customer success and sales and marketing leaders, which fits a company running a negotiated enterprise sale and an open-source community at the same time.

Leadership continuity now runs through the Dynatrace agreement. Dynatrace states that both founders join at closing and that Jason Lopatecki continues to lead the Arize team, reporting directly to its chief executive Rick McConnell, so the people a buyer meets stay in place while the company above them changes. [s4, s9, s10, s12]

Arize publishes the certifications a regulated buyer asks for. The trust center lists SOC 2, PCI DSS, ISO/IEC 27001, GDPR and HIPAA, and the homepage repeats the same list, which answers the procurement question a buyer raises when a vendor inspects proprietary AI traffic.

Arize also lets a buyer choose where the data sits. The pricing comparison puts self-hosted deployment, enterprise single sign-on, audit logs and HIPAA on the enterprise plan and marks service-level agreements standard on the Pro plan and custom on the enterprise one, while a US, EU or CA data region is on every tier and Phoenix runs on a local machine, in a container or in the cloud.

What the probed surfaces do not carry is the paperwork behind the certifications. The trust center names SOC 2, PCI DSS, ISO/IEC 27001, GDPR and HIPAA without publishing the underlying reports, and no federal authorization appears on it, so a procurement team would still request the audit documents itself. [s1, s5, s6]

Add analyzed competitors to compare them side by side with Arize AI.

Arize sells to the AI engineering team at organizations running agents and language-model applications in production. The homepage names chatbots, retrieval systems, copilots and agents as what those teams build, and says they use Arize to understand how those systems behave and to improve them.

The segment runs from a builder on the free tier to a federal program office. The pricing page opens with a free tier for builders and a fifty-dollar tier for AI-native teams before a custom enterprise plan, while the federal spending record carries two Department of Defense purchase orders placed with the same company. That range puts self-serve developers in the same buyer set as a federal program office.

The security buyer is a narrower entry point inside that segment. Arize AX offers two guard types that intercept a message failing a guard, with the default one drawing its examples from a public set of jailbreak prompts, but Arize presents the platform as an engineering tool for teams improving AI quality, so the security need is one capability inside it.

The platform records agent behavior, scores it, and acts on it at run-time. Arize traces each step an agent takes using OpenTelemetry-based instrumentation, runs evaluations over those traces, and integrates with more than forty models, frameworks and AI tools across the major clouds and self-hosted environments.

Phoenix is the capability a buyer can read rather than take on trust. The repository documents tracing, evaluation, datasets, experiments, a playground and prompt management, describes itself as vendor and language agnostic with out-of-the-box support for frameworks including LangGraph, CrewAI and LlamaIndex, and carries 11,174 stars and 1,074 forks.

The guardrails run on prompts and datasets the customer supplies or takes off the shelf. The documentation states that a customer can instantiate a guard with off-the-shelf prompts and datasets from Arize AX or supply its own, and that any model provider can sit behind a guard. A guest column by Jason Bloomberg of Intellyx describes the production shape as continuous lightweight monitoring with judge-model evaluation reserved for high-risk interactions.

Free adoption is the entry point and the paid platform is the destination, and the conversion between them is not disclosed. Arize's own pricing page calls Phoenix the open-source, local-first platform and tells a reader to move to Arize AX only when they need it. That is a stated funnel, not a measured one, and co-founder Aparna Dhinakaran told TechCrunch that the open-source product may be Arize's biggest competitor.

The commercial record contains named accounts and investor backing. TechCrunch reports enterprises including Uber, Klaviyo and Tripadvisor, and names Datadog and PagerDuty among the strategic backers of the 70 million dollar Series C. The customers page adds an LG U+ case study whose named team lead describes evaluation-driven development for 30 million subscribers.

Two federal purchase orders are the traction a reader can verify without the vendor. USAspending records 1,196,635 dollars placed by the Air Force Research Laboratory for language-model evaluation and development tooling, on an order running from July 2025 to July 2026, and a 451,939 dollar order for AI machine learning observability placed by the Navy's NIWC Pacific in 2024 and funded by the Department of the Air Force. No reviewed source states revenue, customer count, or what share of free users pay.

Arize publishes a rate card sized by how much a team observes. The pricing page lists a free tier at 25,000 trace spans and 1 GB of ingestion a month with 15-day retention, a fifty-dollar tier at 50,000 spans and 10 GB with 30-day retention, and custom volumes on the enterprise plan, with startup pricing available on application. Plan sizing tracks instrumentation volume rather than seats, and both published tiers carry unlimited users and unlimited evaluations.

A comparison table sits below the cards and names what the plans differ on. Its security block lists self-hosted deployment, enterprise single sign-on, audit logs, SOC 2 Type II, GDPR, HIPAA and a US, EU or CA data region, which are the terms an enterprise review reads before it asks for a quote.

Published figures stop below the enterprise plan. A large account negotiates a price the public page does not disclose, while the span and volume tiers beneath it give a smaller team a cost it can forecast.

Arize ships as a hosted service with a self-hosted option and a fully self-hostable open-source core. The homepage states that the platform runs across Google Cloud, AWS, Azure and self-hosted environments and that a customer's data stays under its own control, and Phoenix runs locally or self-hosted. A team that will not send traces to a vendor cloud has a supported path.

The operational core is trace ingestion and evaluation at production volume. The platform captures spans from agent and model calls, runs evaluations over them, and applies guardrails at run-time that trigger default responses, retries or blocking. That ingestion load scales with how much a team instruments.

Availability and residency commitments are named rather than specified. The pricing comparison lists service-level agreements and a US, EU or CA data region among the plan features, but the reviewed pages give no numeric availability target, which is the detail an enterprise review asks for before it commits production traffic.

Arize publishes a certification list and describes the programme behind it. The trust center names SOC 2, PCI DSS, ISO/IEC 27001, GDPR and HIPAA compliance, and says the company rests trust on three core pillars. The homepage repeats the same certification list beside the deployment options.

Deployment choice reinforces it. Arize offers a hosted service, a self-hosted option and the fully open-source Phoenix, which runs on a local machine, in a container or in the cloud, and the pricing comparison puts HIPAA, audit logs and enterprise single sign-on on the enterprise plan, with a US, EU or CA data region on every tier.

The record is thinner on the paperwork behind those certifications. The probed trust center lists them without publishing the underlying reports, and no federal authorization appears on it, so a procurement team would still ask for the SOC 2 and ISO documents directly.

Arize positions itself as the layer that AI applications send their traces to, built on open standards. The platform is built on OpenInference and OpenTelemetry, and Phoenix is vendor and language agnostic with out-of-the-box support for frameworks including LangGraph, CrewAI and LlamaIndex, so a buyer is not tied to one cloud's native tooling.

Phoenix extends the ecosystem outward. Phoenix carries 11,174 stars and 1,074 forks and runs locally or self-hosted, and Constellation Research describes a strong open-source community working across all major AI frameworks. Arize AX is the managed AI engineering platform built on those same open standards.

Other vendors can hold the same traces, which is what exposes the position. Datadog sells its own agent observability product and could offer the same recording and scoring inside its own platform, and Dynatrace has agreed to buy Arize outright, which would place the layer inside a general observability platform rather than beside one once the transaction closes.

One founder had led a machine-learning startup that Arize absorbed at launch, and the other came from a company Adobe bought. Aparna Dhinakaran built machine learning infrastructure at Uber and was chief executive of Monitor ML, a Y Combinator-backed startup Arize acquired at launch in 2020. Jason Lopatecki came from TubeMogul, the video advertising company Adobe bought for over 500 million dollars in 2016.

The leadership bench matches an enterprise motion. The about page lists a chief technology officer, a chief information security officer, a compliance officer, a head of customer success and sales and marketing leaders, which is the shape a company needs to run a negotiated enterprise sale and an open-source community at once.

Continuity now runs through the Dynatrace agreement. Dynatrace states that both founders join at closing, that Jason Lopatecki continues to lead the Arize team, and that he reports directly to its chief executive Rick McConnell, so the people a buyer deals with stay in place while the company above them changes.