Arize Phoenix
The fastest self-hosted eval platform to stand up — one pip install, one process, Postgres when you need it — and the best RAG evaluation in the category, but the server is Elastic License 2.0, not open source, and "fully open source, no feature gates" is the vendor's own overstatement.
PRICING
| Phoenix self-hosted | $0 | Elastic License 2.0. No usage caps, no feature gates on features. Tracing, evals, datasets, experiments, prompt management, PXI and the MCP server all included. Runs from pip install arize-phoenix, a single Docker container, Helm, Kustomize or one-click cloud templates. SQLite by default, PostgreSQL 14+ for production. Can run fully air-gapped |
| Phoenix Cloud | $0 | Two hosted Phoenix instances free, no infrastructure setup. The managed path if you do not want to run the container yourself |
| Arize AX Free | $0 | Separate product, not Phoenix. 25,000 spans/month, 1 GB ingestion, 15-day retention, unlimited seats and evals, Alyx assistant, US/EU/CA data region, community support. No overage — you stop collecting at the cap |
| Arize AX Pro | $50/mo | 50,000 spans/month, 10 GB, 30-day retention, unlimited seats, email support, startup discount programme. Hard caps, not overage billing |
| Arize AX Enterprise | Quote only | Custom volumes and retention, SaaS or self-hosted, SOC 2 Type II, HIPAA, GDPR, enterprise SSO, audit logs, RBAC across multiple orgs and spaces, uptime and support SLAs, Signal issue detection, agent swarms, ADB Data Fabric |
Two traps here, both created by the vendor's own naming. First, Phoenix and Arize AX are different products with different SDKs, different authentication and different endpoints — Arize's own documentation opens with an all-caps instruction to coding agents to stop and confirm which one you mean before writing any code. There is no seamless upgrade from Phoenix to AX; you re-instrument. Second, "fully open source, no feature gates" is true about the features and false about the licence. The server is Elastic License 2.0, which is source-available: you can read it, run it, modify it and self-host it, but you may not offer it to third parties as a hosted or managed service. For internal use that restriction never bites. If OSI-approved open source is a procurement requirement, it does. Euro figures converted from published USD at roughly 0.92.
checked 2026-08-15 · vendor pricing page
Element scores
Strengths
Phoenix is the cheapest platform in this category to try and the deepest one to evaluate with. pip install arize-phoenix and phoenix serve gives you a working trace UI in under a minute, and the same binary becomes production-durable by pointing PHOENIX_SQL_DATABASE_URL at Postgres 14+. There is no ClickHouse, no Redis, no separate worker, no object store to provision — one process where the nearest open-source competitor needs four services. For a small team without a platform engineer, that difference is the whole decision. The evaluation surface is the best in the category and independent reviewers keep saying so, particularly on RAG: pre-built judges for hallucination, relevance, toxicity and retrieval quality, code evaluators, dataset evaluators that run as automated test cases, annotation and labeling queues, and every evaluator run traced to its own project so you can audit the judge rather than trust it. Instrumentation is OpenTelemetry-native through OpenInference with coverage across Python, TypeScript and Java and auto-instrumentation for the frameworks people actually build on, which means swapping Phoenix out later is a config change. The agent story is unusually concrete: PXI investigates failing traces, edits prompts as approved diffs and runs experiments, in the browser or from the terminal, and the MCP server is built into the Phoenix process at /mcp with OAuth and per-user scoping, so your coding agent can drive the tool rather than only look at it. Release cadence is genuinely fast — version 20.2.1 shipped on 14/08/2026, with commits landing daily.
Honest dings
Start with the licence, because the marketing does not. The Phoenix server is Elastic License 2.0: source-available, not OSI-approved open source, and it forbids offering Phoenix to third parties as a hosted or managed service. Only the client and eval subpackages are Apache-2.0. "Fully open source, no feature gates" is accurate about feature gating and misleading about freedom, and that distinction survives legal review badly. Second, ingest under load. Issue #12358 is open at priority high and describes the BulkInserter checking out up to seven database connections every 100ms against SQLAlchemy's untuned default pool of fifteen; with two replicas under sustained ingest the pool exhausts and queries fail. Users on the vendor's own community forum report traces surfacing hours after the run at high span volumes. Phoenix started life as an analysis tool and the always-on production deployment is the part you should load-test before you trust it. Third, the product split. Phoenix and Arize AX share a brand, a documentation domain and a search index, but not their SDKs — Arize's docs literally shout at coding agents to confirm which product is meant before writing code, and there is no migration path from one to the other. Anyone who adopts Phoenix expecting a smooth graduation to AX will re-instrument. Fourth, the things that are AX-only and matter in production: online evals against live traffic, production monitors, alerting, SOC 2 and HIPAA coverage, and any uptime commitment. Self-hosted Phoenix has none of those, and the open-issue count sits above 900.
Sources (21) — every claim traceable
Every audit lists the research it rests on — transparency and traceability are the product. Tools evolve: each audit is a snapshot of its audit date, and re-audits supersede older versions (kept below for reference).
- arize.com/docs/phoenix — Official product documentation: tracing on OpenTelemetry via OpenInference, evaluation (LLM-based, dataset evaluators, human annotations, Ragas/DeepEval/Cleanlab integrations), prompt management with versioning and span replay, datasets and experiments, PXI and the remote MCP server. Source for the feature scope behind Pr, Cx, Tr, Ev (accessed 2026-08-15)
- github.com/Arize-ai/phoenix/blob/main/LICENSE — Primary evidence for the licence finding: the server repository ships the full Elastic License 2.0 text, including the clause forbidding provision of the software to third parties as a hosted or managed service. This is what makes 'fully open source' an overstatement (accessed 2026-08-15)
- arize.com/docs/phoenix/self-hosting — Official self-hosting docs: free to self-host with no licence fees, usage limits or feature gates, data never leaves your infrastructure, air-gapped deployment supported. Deployment matrix across CLI, Docker, Kubernetes, Helm, CloudFormation and one-click cloud templates; Docker Hub tags including nonroot and debug variants (accessed 2026-08-15)
- arize.com/docs/phoenix/self-hosting/deployment-opti… — Official Docker deployment guide and the evidence for the single-process claim: one container serving UI plus OTLP HTTP on 6006 and OTLP gRPC on 4317, SQLite by default, PostgreSQL 14+ minimum for production via PHOENIX_SQL_DATABASE_URL (accessed 2026-08-15)
- arize.com/docs/phoenix/integrations/remote-mcp — Official MCP documentation and the basis for the Protocols score: MCP server built into the Phoenix server at /mcp from version 19.0.0, OAuth authorization code plus PKCE with dynamic client registration, tokens scoped to user permissions, API-key bearer fallback for headless environments, px setup mcp writing configs for six clients. Vendor states the server is beta and enabled by default (accessed 2026-08-15)
- arize.com/docs/phoenix/pxi — Official PXI documentation: server-side engineering agent in browser and terminal, requires your own provider credentials, defaults to claude-opus-5, capability flags for subagents, bash, web access and GraphQL mutations, edits shown as approved diffs. Vendor's own beta warning that it can and will make mistakes (accessed 2026-08-15)
- arize.com/docs/phoenix/evaluation/llm-evals — Official evaluation docs: model-agnostic judge adapters, pre-built RAG and tool-calling metrics, structured output enforced through tool calling rather than text parsing, executors handling rate limits and dynamic concurrency, every evaluator run traced to a dedicated project, explanations returned by default. Also the explicit note that online evals on production traffic are an Arize AX feature (accessed 2026-08-15)
- arize.com/docs/phoenix/inferences/phoenix-inferences — Official Inferences docs and the evidence for the Embeddings score: arize-phoenix[embeddings] extra, UMAP projection, automatic HDBSCAN clustering, training-versus-production drift comparison, corpus set for retrieval analysis (accessed 2026-08-15)
- arize.com/docs/phoenix/self-hosting/features/authen… — Official auth configuration: PHOENIX_ENABLE_AUTH and PHOENIX_SECRET, OAuth2, LDAP and local accounts, RBAC, brute-force lockout after five attempts, optional strong password policy, CSRF trusted origins. Also the warning that enabling auth stops trace collection until API keys exist (accessed 2026-08-15)
- arize.com/docs/phoenix/settings/data-retention — Official retention docs: indefinite retention by default, per-project policies by age or trace count on a CRON schedule, PHOENIX_DEFAULT_RETENTION_POLICY_DAYS at deploy time, and the pointer that multi-party data governance and auditability route to Arize AX (accessed 2026-08-15)
- arize.com/docs/phoenix/get-started — Primary evidence for the product-split finding: the page opens with an all-caps instruction to AI coding agents to stop and verify whether the user means Phoenix or Arize AX before writing any code, stating the two are completely different products with different APIs, authentication and endpoints (accessed 2026-08-15)
- arize.com/pricing — Official pricing verified live 15/08/2026: AX Free $0 at 25k spans, 1 GB and 15 days; AX Pro $50/month at 50k spans, 10 GB and 30 days; AX Enterprise custom. Full feature matrix showing self-hosted deployment, enterprise SSO, audit logs, HIPAA and agent swarms as Enterprise-only, SOC 2 Type II and GDPR from Pro, unlimited seats and evals on every tier, US/EU/CA data regions (accessed 2026-08-15)
- github.com/Arize-ai/phoenix/issues/12358 — Independent user bug report, open and labelled priority high, assigned to a core maintainer: BulkInserter._bulk_insert opens up to seven DB sessions per 100ms iteration against SQLAlchemy defaults of pool_size 5 and max_overflow 10, exhausting the pool under sustained ingest with two replicas. Reporter offered a fix; maintainer replied that the area is sensitive and a core maintainer would handle it (accessed 2026-08-15)
- community.arize.com/x/phoenix-support/0vzh05w5uqmq/… — Independent user report on the vendor's own community forum: traces appearing in the UI hours after the run on a self-hosted GCP deployment with a separate Postgres, sessions above 7,000 spans, gRPC returning resource-exhausted. Arize's head of OSS confirms spans queue in memory to protect against concurrent writes and advises decoupling Phoenix from the application host (accessed 2026-08-15)
- llmtools.cc/tool/arize-phoenix — Independent review making the precise licence distinction corroborated against the LICENSE file: server is ELv2 while phoenix-evals, phoenix-client and phoenix-otel are Apache-2.0, so the eval libraries are open source and the server you self-host is not. Also flags ingest lag tied to the bulk-inserter path (accessed 2026-08-15)
- agenticwire.news/article/langfuse-vs-arize-phoenix — Independent head-to-head (02/07/2026) with a first-hand operator note: Phoenix is a single process on SQLite or Postgres 14+ where Langfuse needs Postgres, ClickHouse, Redis/Valkey and S3-compatible storage plus two application containers. Positions Phoenix as stronger on evaluation and RAG tracing, Langfuse as stronger on production monitoring and prompt governance (accessed 2026-08-15)
- chatforest.com/reviews/helicone-llm-observability-g… — Independent review scoring Phoenix 4/5 and corroborating the single-container advantage over Langfuse's Postgres plus ClickHouse plus Redis stack, the ELv2 constraint as the most significant drawback, the Phoenix/AX product-split confusion, and the roughly 500 open issues at time of writing (accessed 2026-08-15)
- langchain.com/resources/langsmith-vs-arize — Competitor comparison from LangChain, used only for its checkable factual claims: Phoenix and Arize AX ship incompatible SDKs so there is no seamless migration from the free tool to the paid platform, CI/CD quality gates need manual integration, and it cites Phoenix issues #12358, #11873 and #12349. Treated as an interested source and cross-checked against the Arize docs and GitHub, both of which corroborate (accessed 2026-08-15)
- github.com/Arize-ai/phoenix — Official repository measured 15/08/2026: 11,062 stars, 1,058 forks, 922 open issues, licence reported as Other (NOASSERTION), created November 2022. Release arize-phoenix-v20.2.1 published 14/08/2026 with commits landing daily, which is the evidence for the release-cadence claim (accessed 2026-08-15)
- prnewswire.com/news-releases/arize-ai-secures-70m-s… — Independent confirmation of the $70M Series C announced 21/02/2025, led by Adams Street Partners with M12, Sinewave, OMERS Ventures, Datadog, PagerDuty, Industry Ventures and Archerman participating, and existing investors Foundation Capital, Battery Ventures, TCV and Swift Ventures. Company based in Berkeley, California (accessed 2026-08-15)
- arize.com/about-us — Official company page: founders Jason Lopatecki, Aparna Dhinakaran and Michael Schiff, offices in Berkeley and New York, named CISO, 26 open positions at time of access (accessed 2026-08-15)