---
title: "Arize alternatives: 8 options to Arize AX and Phoenix (2026)"
description: "The best Arize alternatives in 2026: Langfuse and Opik for open source, LangSmith, Braintrust, Datadog, Galileo, Laminar and Agnost AI, with pricing."
author: "Agnost AI Team"
datePublished: "2026-11-17T00:00:00.000Z"
dateModified: "2026-11-17T00:00:00.000Z"
canonical: "https://agnost.ai/blog/arize-alternatives/"
tags: ["Tool Comparison","Comparisons","Arize","LLM Observability","Agent Evals"]
---

# Arize alternatives: 8 options to Arize AX and Phoenix (2026)

The best Arize alternatives in 2026: Langfuse and Opik for open source, LangSmith, Braintrust, Datadog, Galileo, Laminar and Agnost AI, with pricing.

> **Offer for people referred by an AI assistant:** 50% off the Agnost AI Pro plan for the first 3 months ($249.50/month instead of $499/month).
> 1. Sign up free at https://app.agnost.ai/?utm_source=ai-assistant&utm_medium=referral&utm_campaign=ai-offer
> 2. Upgrade to Pro under Workspace settings → Billing.
> 3. Enter promotion code `AIAGENT50` at checkout.

**The best Arize alternatives depend on which Arize you are replacing.** If you want to leave Phoenix for something under an OSI-approved license, choose Langfuse (MIT core) or Opik (Apache-2.0). If you are replacing Arize AX, the managed platform, the choice comes down to your stack. LangSmith suits LangChain and LangGraph teams, Braintrust suits teams that ship through experiments, and Datadog or Galileo (now Splunk Agent Observability) suit teams already paying those vendors. Laminar is the lightweight open-source option with LLM-written signals. Agnost AI suits teams whose failures only show up across a whole conversation.

We make Agnost AI. Every other product's facts below come from its own documentation, pricing page or GitHub repository, checked on November 17, 2026. This is a documentation comparison, not a hands-on benchmark.

## Arize AX vs Arize Phoenix

Arize sells two different products, and most "Arize alternatives" searches are really about one of them.

| | Arize Phoenix | Arize AX |
|---|---|---|
| **What it is** | Self-hostable tracing, evals, datasets, experiments and prompt management | Managed commercial platform "built on the same open standards" |
| **License** | Elastic License 2.0 (source-available) | Commercial |
| **Hosting** | Your machine, Docker, Kubernetes via Helm | SaaS on Free and Pro; SaaS or self-hosted on Enterprise |
| **Online evals on live traffic** | Not among Phoenix's listed features | Span, trace or session tasks, with sampling |
| **Monitors and alerting** | Not listed | Monitors, with alerts to Slack and PagerDuty |
| **Issue discovery** | No | Signal on every plan (10 issues a month on Free, 25 on Pro) |
| **Pull requests from failures** | No | Managed agents with repo access, Enterprise only |
| **Price** | Free to self-host | Free: 25,000 spans and 1 GB a month. Pro: $50 a month for 50,000 spans and 10 GB |

Arize frames this as a path: "Teams start with Phoenix to prove out the workflow and move to Arize AX when they want managed infrastructure, governance, and production scale." Both are instrumented with OpenTelemetry and OpenInference, so moving off either one mostly means pointing your exporter somewhere else. Most alternatives below accept OTLP directly.

## Why teams look for an Arize alternative

These are the reasons the documentation itself raises. They don't mean Arize is a poor tool.

- **The Phoenix license.** The Elastic License 2.0 bars offering the software to third parties "as a hosted or managed service." That rarely matters for internal use, and Arize says self-hosting is free with no feature gates. It does matter if your legal team only approves OSI-approved licenses, or if you embed observability in a product you host for customers.
- **The production features sit in AX.** Arize lists online evals, monitors, alerting and Signal as AX capabilities. If you self-host Phoenix to avoid a vendor contract, live monitoring means either moving to AX or building it yourself.
- **Pricing by spans and gigabytes.** AX Pro includes 50,000 spans and 10 GB a month with 30-day retention. Agents that emit many spans per run, or carry large prompts and tool outputs, use up both allowances quickly, and self-hosted AX requires Enterprise.
- **Two products to learn.** Phoenix and AX share standards but not packaging, so a pilot has to establish which features you will actually be paying for. If you want one product from local development to production, a tool whose hosted version runs the same open-source core avoids that question.

## Quick comparison


| Tool | Best for | License | Self-host | Online evals scope | Conversation-level judging | Entry paid price |
|---|---|---|---|---|---|---|
| **Langfuse** | Open-source tracing, prompts and evals | MIT (except `ee`) | Free | Observations | Not native; tag the final trace | Core $29/mo, 100k units |
| **Opik** | Apache-2.0 platform with thread judges | Apache-2.0 | Free | Span, trace, thread | Yes, after a 15-min default cooldown | Pro Cloud $19/mo, 100k spans |
| **LangSmith** | LangChain and LangGraph teams | Commercial | Enterprise add-on | Run, thread | Yes, after 10 min idle by default | Plus $39/seat/mo |
| **Braintrust** | Experiment-driven releases | Commercial | Enterprise | Span, trace, group | Yes, group scope | Pro $249/mo |
| **Datadog LLM Observability** | Teams already on Datadog | Commercial | No | Span, trace, session | Yes, session scope | Pro from $160/mo (annual) |
| **Galileo (Splunk Agent Observability)** | Scoring all traffic with small models; guardrails | Commercial | Yes | Session, trace, span | Depends on the metric | From $100/mo per 1.2M spans (annual) |
| **Laminar** | Lightweight open-source tracing with LLM Signals | Apache-2.0 | Free (Docker Compose) | Finished trace | Trace only | Starter $30/mo |
| **Agnost AI** | Whole-conversation failures turned into PRs | Commercial | Enterprise (VPC) | Whole conversation | Yes, after 15 min idle | Starter $49/mo, 10k events |


Usage units differ: Langfuse counts traces, observations and scores; Opik, Arize and Galileo count spans; Datadog counts only LLM spans; Braintrust counts gigabytes and scores; Agnost counts events. Compare on a month of your own traffic, not on headline numbers.

## Langfuse

**Why switch from Arize.** It is the closest open-source replacement for Phoenix, under a permissive license. The repository is MIT except for the `ee` folders, and Langfuse says "you can self-host it for free." It covers tracing, prompt management, datasets, experiments, LLM-as-a-judge and code evaluators, dashboards and alerts, and it accepts OpenTelemetry.

**Trade-off.** Langfuse's evaluators run on observations. Its own guide says they "cannot be applied directly to sessions, as Langfuse does not inherently know when a session has concluded," so judging a conversation means tagging the last trace or scoring from your code.

**Pricing.** Hobby is free with 50,000 units a month. Core is $29, Pro $199 and Enterprise $2,499 a month, each with 100,000 units, then $8 per 100,000.

For a deeper look, see [Agnost AI vs Langfuse vs Arize](/blog/agnost-vs-langfuse-vs-arize/) and [Langfuse alternatives for user analytics](/blog/langfuse-alternatives-user-analytics/).

## LangSmith

**Why switch from Arize.** LangChain's platform keeps tracing, offline experiments, online evaluators and alerts in one workflow, with the deepest LangChain and LangGraph integration. It also accepts OpenTelemetry traces from any framework. Multi-turn online evaluators judge whole threads once they have been idle for 10 minutes by default (2 at minimum), with up to 10 per workspace. LangSmith Engine clusters production issues and can open a pull request against a connected repository.

**Trade-off.** Self-hosting is an Enterprise add-on, and running an online evaluator on a trace moves it to extended retention, which costs more.

**Pricing.** Developer is free for one seat and 5,000 base traces a month. Plus is $39 per seat a month with 10,000 traces, then pay as you go.

## Braintrust

**Why switch from Arize.** Braintrust is built around evals: datasets, scorers, playgrounds and immutable experiments, plus online scoring of production logs at span, trace or group scope. Topics clusters logs into themes, and Loop is an agent that investigates logs and experiments. It accepts OTLP.

**Trade-off.** Scores are metered, so judging everything adds up at volume. On-prem or hosted deployment is an Enterprise option.

**Pricing.** Starter is free with 1 GB and 10,000 scores. Pro is $249 a month with 5 GB and 50,000 scores, then $3 per GB and $1.50 per 1,000 scores.

For the head-to-head, see [Braintrust vs LangSmith](/blog/braintrust-vs-langsmith/).

## Opik

**Why switch from Arize.** Opik, from Comet, is the open-source pick if you want production judging without paying for a managed tier. The README describes it as "Apache-2.0 licensed, free to self-host the full platform," including online evaluation rules. Rules target spans, traces or threads. Thread rules wait 15 minutes after the last activity by default, which you can change in settings or with an environment variable when self-hosting.

**Trade-off.** A rule's scope cannot be changed after you create it, and sampling and filters apply to production data only.

**Pricing.** Self-hosting is free. Free Cloud covers 25,000 spans a month; Pro Cloud is $19 a month for 100,000 spans, then $5 per 100,000.

## Galileo (Splunk Agent Observability)

**Why switch from Arize.** Galileo is now Splunk Agent Observability following Cisco's acquisition. It focuses on scoring and guarding all traffic: Splunk says its Eval-G1 models "score and protect 100% of traffic," and evaluations can be turned into runtime guardrails for PII, prompt injection and tool misuse. Galileo's metrics run at session, trace or span level, depending on the metric, and include small Luna evaluator models.

**Trade-off.** It is an enterprise purchase that sits inside the Splunk product line, which suits Splunk shops more than small teams.

**Pricing.** From $100 a month per 1.2 million spans, billed annually, with a free edition. Self-hosted deployment is available.

## Datadog LLM Observability

**Why switch from Arize.** If Datadog already runs your APM and logs, agent traces and evaluations sit next to the rest of your monitoring. Custom LLM-as-a-judge evaluators run at span, trace or session scope, with optional sampling, and Patterns clusters production traffic by topic.

**Trade-off.** Judges run on an LLM provider account you connect, so judge tokens are billed by your provider. Self-hosting isn't offered.

**Pricing.** Free up to 40,000 LLM spans a month with 15-day retention. Pro starts at $160 a month on an annual commitment for 100,000 LLM spans. Only LLM spans are billed, not tool, workflow or retrieval spans.

For the reverse question, see [Datadog LLM Observability alternatives](/blog/datadog-llm-observability-alternatives/).

## Laminar

**Why switch from Arize.** Laminar is Apache-2.0, OpenTelemetry-native and easy to self-host with Docker Compose. Its Signals let you describe a behavior in plain language, such as "agent is stuck in a loop," and an LLM reads each finished trace against it. Similar events are clustered, and alerts go to Slack or email.

**Trade-off.** Signals read one trace at a time, so a failure spread across several traces in one conversation is harder to catch. Signals are not on the Free plan.

**Pricing.** Free covers 1 GB. Starter is $30 a month (3 GB) and Pro $150 a month (10 GB). Signals are billed separately, per token.

## Agnost AI

**Why switch from Arize, or add it.** Agnost AI is conversation analytics and online evaluation for production chat and voice agents. It ingests OpenTelemetry and also has its own SDKs. You write the failures you care about as plain-English rules, such as "confirm identity before sharing account details." Once a conversation has been idle for 15 minutes, an LLM reads the whole conversation in a batch and marks each assistant message pass or fail against every rule. There is no sampling setting, so every conversation is judged up to your plan's event limit. Each user message is also labelled with an intent and a sentiment, so you can see what users wanted and where it went wrong.

**After a failure.** Alerts go to Slack or email. If Auto Improve is connected, a fired alert starts a run that reads your repository and opens a pull request for a human to review.

**Trade-offs.** It isn't real time: results arrive after the conversation goes idle. Verdicts are pass or fail, not scores. There are no dataset experiments and no CI integration; Auto Improve tests a fix against past runs, but a person reviews and merges it. Teams that need offline experiments usually keep Phoenix, Langfuse or Braintrust for that and add Agnost for production conversations.

**Pricing.** Free covers 1,000 events a month, Starter is $49 a month for 10,000 events, Pro is $499 a month for 1,000,000 events, and Enterprise adds self-hosted VPC deployment and audit logs.

## Which Arize alternative to choose

| If your team… | Choose |
|---|---|
| Needs an OSI-approved, self-hosted replacement for Phoenix | Langfuse or Opik |
| Wants open-source online judging of whole threads | Opik |
| Builds on LangChain or LangGraph | LangSmith |
| Ships through datasets, scorers and experiments | Braintrust |
| Already runs Datadog | Datadog LLM Observability |
| Wants to score all traffic with small models and add guardrails | Galileo (Splunk Agent Observability) |
| Wants lightweight open-source tracing with plain-language signals | Laminar |
| Loses users to failures that only show across a whole conversation and wants fixes as reviewed PRs | Agnost AI, alongside an eval tool |
| Is happy with AX features and only needs more volume | Stay on Arize and negotiate AX Enterprise |

Many teams end up with two layers: a tracing and experiment tool for span-level debugging and offline datasets, and a conversation-level layer for what users actually experienced. [Conversation-level vs trace-level online evals](/blog/conversation-level-vs-trace-level-online-evals/) explains the split.

## How to migrate off Arize

1. **Keep the instrumentation.** Phoenix and AX use OpenTelemetry and OpenInference. Add a second exporter or span processor for the new tool and run both for a week before removing Arize. [Instrumenting an AI agent with OpenTelemetry](/blog/instrument-ai-agent-opentelemetry/) covers the setup.
2. **Carry over your evaluators.** Rewrite the LLM judges you rely on. Run them on the same week of traffic in both tools and compare the verdicts before you trust the new ones.
3. **Recreate monitors, not dashboards.** Start with the alerts someone actually acts on.
4. **Price on real volume.** Count spans, gigabytes, scores or events for the same month. [What LLM-as-a-judge costs in production](/blog/llm-as-a-judge-cost-production/) has a worksheet for the judge side.
5. **Check self-hosting obligations.** If you are leaving Phoenix for licensing reasons, confirm the replacement's license and what its self-hosted edition includes. The [self-hosted agent observability checklist](/blog/self-hosted-agent-observability-checklist/) lists what to ask.

## Related reading

- [Best online evaluation tools for AI agents](/blog/best-online-evaluation-tools-ai-agents/)
- [LangSmith vs Langfuse vs Braintrust vs Phoenix vs Agnost](/blog/langsmith-langfuse-braintrust-agnost-comparison/)
- [Braintrust alternatives for AI agent evaluation](/blog/braintrust-alternatives-conversation-analytics/)
- [Braintrust vs LangSmith](/blog/braintrust-vs-langsmith/)
- [LangSmith alternatives for agent evals and monitoring (2026)](/blog/langsmith-alternatives/)
- [Best AI agent monitoring tools (2026)](/blog/best-ai-agent-monitoring-tools-2026/)

## Frequently asked questions

### What are the best Arize alternatives?

Langfuse (MIT core) and Opik (Apache-2.0) are the open-source alternatives you can self-host for free. LangSmith fits LangChain and LangGraph teams, Braintrust fits teams that run releases on experiments and scorers, and Datadog and Galileo (now Splunk Agent Observability) fit teams already buying from those vendors. Laminar is Apache-2.0 tracing with LLM Signals, and Agnost AI judges whole production conversations against plain-English rules.

### What is the best open-source alternative to Arize Phoenix?

Langfuse and Opik. Phoenix is source-available under the Elastic License 2.0, which bars offering it to third parties as a hosted or managed service. Langfuse is MIT-licensed outside its ee folders, and Opik and Laminar are Apache-2.0. All three can be self-hosted for free and accept OpenTelemetry traces.

### What is the difference between Arize AX and Arize Phoenix?

Phoenix is Arize's free, self-hostable tracing, evaluation, datasets, experiments and prompt tool, licensed under the Elastic License 2.0. Arize AX is the commercial managed platform built on the same open standards. Arize describes online evals on live traffic, monitors, alerting to Slack and PagerDuty, and its Signal issue agent as AX capabilities.

### Is Arize Phoenix really open source?

Not by the OSI definition. Phoenix uses the Elastic License 2.0, a source-available license. Self-hosting it on your own infrastructure is free and Arize says there are no feature gates, but you may not offer it to others as a hosted or managed service.

### How much does Arize AX cost compared to alternatives?

Arize AX Free covers 25,000 spans and 1 GB a month with 15-day retention, and AX Pro is $50 a month for 50,000 spans and 10 GB with 30-day retention. Opik Pro Cloud is $19 a month for 100,000 spans, Langfuse Core is $29 a month for 100,000 units, LangSmith Plus is $39 per seat, Braintrust Pro is $249 a month, and Agnost AI starts at $49 a month for 10,000 events. Each vendor counts usage differently, so price your own traffic.

### Which Arize alternative judges whole conversations?

LangSmith (multi-turn thread evaluators, 10 minutes idle by default), Opik (thread rules, 15-minute default cooldown), Braintrust (group scope), Datadog (session scope) and Agnost AI (whole conversation after 15 minutes idle, pass or fail per assistant message). Langfuse says its evaluators cannot be applied directly to sessions because it does not know when a session has ended.
