Arize Phoenix

💰 Pricing & Plans

🔧 API & Developer Experience

• •OpenTelemetry Native: Phoenix is an OpenTelemetry-native trace backend; you can throw OTLP traces from any instrumented app straight at it without vendor-specific SDKs.
• •Auto-instrumentation: 1-line wrappers or environment variables auto-capture traces for 17+ frameworks: OpenAI, Anthropic, LangChain, LlamaIndex, vLLM, Ollama, DSPy, and more.
• •Traces, Spans & Evaluations: Deep span/trace explorer for LLM calls, plus first-class evaluation tooling (LLM-as-judge, custom scorers, human review) that runs against captured traces.
• •Prompt Playground: Iterate prompts and compare model outputs live; experiment tracking records prompt versions alongside evaluation scores.
• •PXI Agent (BETA): Phoenix Intelligence lets you ask natural-language questions about your traces ('where are failures spiking?') and get investigation answers grounded in your data.
• •Datasets & Experiment Hub: Versioned datasets, regression testing, and experiment comparison built into the platform — closest to a CI/CD loop for prompt and RAG changes.
• •Open Licenses & Local UI: Self-host ships Apache-2/Elastic-2.0 licensed server + local web UI (no telemetry), so raw trace data never leaves your VPC.

🧩 OpenTelemetry-Native LLM Observability

Arize Phoenix stands out because it is one of the few LLM observability platforms built directly on the OpenTelemetry standard rather than a proprietary instrumentation format. Teams already emitting OTLP traces can feed Phoenix immediately — or use one-line auto-instrumentation for 17+ frameworks — without re-architecting their stack or adopting a vendor lock-in SDK. This means the same trace pipeline that serves your general microservices can also power deep LLM-specific analysis: span trees, token/cost per call, prompt–completion pairs, and retrieval quality for RAG pipelines. Phoenix folds evaluation directly into the observability loop. Instead of exporting traces to one tool and running evals in another, you capture traces, score them with LLM-as-judge or custom scorers, log datasets, and iterate prompts in one place. The optional PXI agent adds natural-language investigation over the whole corpus. Combined with an Elastic-2.0 open-source self-host option, it delivers production-grade tracing, evals, and experiment tracking with minimal vendor lock-in — a strong fit for platform teams building internal LLM infrastructure.

🌐 China Access & Latency

Arize Phoenix has a clear advantage for China-based teams: the core product is open source and self-hostable, so mainland teams can run the Phoenix server on a domestic cloud or on-premises without proxying to any overseas SaaS. The self-hosted server ships an Apache-2/Elastic-2.0 licensed local web UI with no external telemetry, and trace data stays inside the VPC — ideal for compliance-sensitive workloads subject to Chinese data-residency rules. Because it is an observability backend rather than a hosted model API, there is no per-token egress cost and no China-specific rate-limiting at the provider edge. The trade-off appears only if you use Phoenix Cloud (the managed SaaS hosted overseas): reaching it from mainland China then requires a VPN or an overseas relay, with typical added latency in the hundreds of milliseconds. For teams that want the managed experience with a Chinese-accessible path, self-hosting Phoenix on a domestic cloud and connecting clients locally is the standard, effectively latency-free approach.