Arize Phoenix
Choose Phoenix for operating your own AI trace and experiment store, after checking the ELv2 service restriction.
- Starting price
- ELv2 self-hosted software; operating costs separate
- Free access
- Self-hosted software without a listed subscription fee
- Billing unit
- ELv2 software; compute, storage and evaluator calls
- Platform
- Python server; Arize maintainer mark shown
- Verified
- 1 Nov 2026
Overview
What you provide
- Send instrumented application spans and authorised evaluation examples to a Phoenix instance you operate.
What you get
- Phoenix returns trace views, versioned datasets and experiment results for debugging model and retrieval behaviour.
Choosing this tool
The README documents pip install arize-phoenix followed by phoenix serve, or uvx arize-phoenix serve. OpenTelemetry/OpenInference instrumentation sends traces to your instance; datasets, experiments and a prompt playground help investigate failures. LangWatch is a managed event-metered alternative, while promptfoo concentrates on repeatable CLI evaluations. Local hosting changes where traces live, but an external evaluator model can still receive selected trace content.
A decision worked through
A local server with 500 stored traces does not itself imply 500 paid model calls. Running one external judge for every trace would create 500 judge requests under that assumption, billed by the chosen provider.
Arize Phoenix alternatives for self-hosted ai tracing and experiments
Compare Arize Phoenix with LangWatch, promptfoo for this workflow, including the billing unit and the work left to your team.
| Product | Price / unit | Free allowance | Input → output | Key limit |
|---|---|---|---|---|
| Arize PhoenixRead profileOfficial source ↗ | ELv2 self-hosted software; operating costs separate | No listed self-hosted subscription charge | Instrumented spans become traces and experiment datasets. | ELv2 restricts offering substantial features as a managed service. |
| LangWatchRead profileOfficial source ↗ | Growth EUR 29/core seat/month | Developer includes 50,000 monthly events | Agent traces and datasets yield evaluation results. | Calls and evaluation steps can count as separate events. |
| promptfooRead profileOfficial source ↗ | MIT CLI software; inference billed separately | No core CLI licence fee | Test cases and model configurations yield comparative evaluations. | Remote providers can receive prompts despite local orchestration. |
Setup & workflow
- Install arize-phoenix and start the local server using the documented command.
- Instrument a test application and inspect one complete trace.
- Create an evaluation dataset and review model-call costs and licence boundaries.
Requirements & installation
- Prepare a Python environment or container and configure storage, access control and trace redaction for the intended deployment.
Limits & review
- ELv2 is not a blanket permission to resell a managed Phoenix service. Hosting, persistence and external evaluator costs remain the operator’s responsibility.
Your part
- Inspect trace payloads and evaluator destinations and check the licence before exposing a hosted service to others.
Sources & verification5
This profile is based on official sources, not a hands-on product test.
Vendor descriptions and demos document advertised features. Editorial guidance is based on these sources.
- Official Arize Phoenix pricinggithub.com
- Official Arize Phoenix productarize.com
- Arize Phoenix official referencegithub.com
- LangWatch official referencelangwatch.ai
- promptfoo official referencegithub.com
