| Type | Startup | Startup | Startup |
|---|
| Categories | | | |
|---|
| In one line | Open-source LLM engineering platform: tracing, prompts, evals. | Enterprise-grade evals, data and AI gateway. | Observability and evaluation for AI — makers of OSS Phoenix. |
|---|
| Stage & fundingDiffers | Y Combinator · Acquired by ClickHouse (Jan 2026) | — | — |
|---|
| GitHub stars | — | — | — |
|---|
| License | — | — | — |
|---|
| Language | — | — | — |
|---|
| Last push | — | — | — |
|---|
| Works with | — | — | — |
|---|
| Built on | — | — | — |
|---|
| Built on it | — | — | — |
|---|
| Similar toDiffers | | — | |
|---|
| People linkedDiffers | 2 | 1 | 2 |
|---|
| Events | 0 | 0 | 0 |
|---|
| Pros | - +DocsCore features are MIT-licensed and free to self-host with Docker Compose or Kubernetes, and ClickHouse committed to keeping that license after the acquisition.[1][2]
- +DocsThe free Hobby cloud tier includes 50k units per month, 2 users and 30 days of data access with no credit card.[1]
- +MetadataWidely adopted, with more than 20,000 GitHub stars and over 26 million SDK installs per month at the end of 2025.[1]
- +DocsPro costs $199/month and adds 3 years of data access, data-retention controls, SOC2/ISO27001 reports and an available HIPAA BAA.[1]
| - +User reportsUsers find it easy to set up datasets and run evals across models; one calls it a game changer for checking prompts, especially for classification.[1][2]
- +Independent testTracing, evals, datasets and prompts in one product, which a six-month hands-on review called the most complete eval toolchain in 2026.[1]
- +Independent testYou can turn any production trace into a golden eval example from the UI, and the free tier includes 1M trace spans and 10K scores a month.[1]
- +User reportsPractitioners use it for model comparisons such as long-context retrieval under production constraints.[1]
| - +MetadataPhoenix is free to self-host and very active (about 12k stars, commits on 29 Oct 2026).[1][2]
- +User reportsBuilt on OpenTelemetry and OpenInference, so traces are portable; users praise OpenInference, and a community proxy sends Claude Code spans to Phoenix.[1][2]
- +User reportsPractitioners use Phoenix for LLM-as-judge plus code-based evals and continuous analysis of production traces.[1][2]
- +DocsEvery AX plan, including Free, has unlimited users and evals, and AX Pro costs $50/month.[1]
|
|---|
| Cons | - −Independent testSelf-hosting v3 means running ClickHouse, Redis, blob storage, a worker and Postgres instead of v2's single container, and guides cite about 4 cores and 9GB RAM as the minimum.[1][2]
- −User reportsSelf-hosters report that a Redis outage or out-of-memory error stops event processing entirely, along with recurring ClickHouse migration failures and stuck trace deletion.[1]
- −DocsEnterprise SSO and fine-grained RBAC need the $300/month Teams add-on on top of Pro, and audit logs and SCIM need the $2,499/month Enterprise plan.[1]
- −MetadataClickHouse now owns it and plans tighter integration with its database, so a data-platform vendor sets the long-term roadmap.[1]
| - −Independent testClosed source with no self-hosting on standard plans; only Enterprise gets a VPC option.[1][2]
- −Independent testLock-in risk: datasets, eval configs and trace history live in Braintrust, and the review found the export tooling is not prominent.[1]
- −Independent testLLM-as-judge scorers add model costs on top of the platform fee, estimated at $500-2,000/month for 10K evals a day with expensive judges.[1]
- −User reportsPro costs $249/month, and one happy user said the core features are simple enough that they would build them themselves next time.[1][2]
| - −MetadataPhoenix uses the Elastic License 2.0, which is source-available rather than OSI open source and restricts offering it as a managed service.[1]
- −DocsAX Free caps you at 25k spans a month with 15-day retention, and Pro raises that only to 50k spans and 30 days.[1]
- −DocsSelf-hosted AX, SSO and audit logs are Enterprise-only.[1]
- −User reportsOne user found Arize's Claude Code tracing setup heavyweight, and others dropped Phoenix after reviewing its site.[1][2]
- −MetadataPhoenix has about 1,000 open GitHub issues.[1]
|
|---|
| Best for | Teams that want open-source LLM tracing, prompt management and evals with a real option to self-host. | Teams that want a polished, hosted eval and tracing workflow from day one. | Teams that want OpenTelemetry-native LLM tracing and evals they can run themselves with Phoenix, with a hosted option later. |
|---|
| Watch out for | Self-hosted v3 is a multi-service deployment, so budget ops time or use the cloud tiers. | Plan how you will export datasets, and budget for LLM-judge costs. | AX Free and Pro have low span and retention caps, and ELv2 limits offering Phoenix as a service. |
|---|
| Links | | | |
|---|