Skip to main content
Python classification evaluators can now use a decision model as the judge, and @arizeai/phoenix-otel 3 moves TypeScript tracing to OpenTelemetry JS 3. The px CLI lists prompt and dataset versions and tags prompt versions.

Breaking Change: OpenTelemetry JS 3 in @arizeai/phoenix-otel

November 11 to 16, 2026 Breaking change in @arizeai/phoenix-otel 3.0.0 and @arizeai/phoenix-client 8.0.0; sampler and spanLimits in @arizeai/phoenix-otel 3.1.0+ @arizeai/phoenix-otel now runs on OpenTelemetry JS 3, so spans from apps that use OpenTelemetry JS 3 are recorded after register().
  • Node.js 22.15 or newer and @opentelemetry/api 1.10 or newer are required.
  • Span processors take an options object. Import BatchSpanProcessor and SimpleSpanProcessor from @arizeai/phoenix-otel and pass { exporter }. These processors throw a TypeError for a positional exporter instead of silently dropping every span.
  • register() returns a provider without a register() method. Call attachGlobalTracerProvider(provider) instead. A tracerProvider passed to evaluateExperiment() must also come from OpenTelemetry JS 3.
  • Set sampling and limits in code. OpenTelemetry JS 3 ignores OTEL_TRACES_SAMPLER, OTEL_SPAN_*, OTEL_ATTRIBUTE_*, and OTEL_BSP_*, so register() accepts sampler and spanLimits and logs a warning for each ignored variable.

Migration Guide

Every step for upgrading from v2 to v3

@arizeai/phoenix-otel

Registration options

Decision-Model Evaluators in Python

November 10, 2026 Available in arize-phoenix-evals 3.10.0+ Python classification evaluators can now use a decision model as the judge. A decision model answers a typed choice question directly, so each score carries a label and per-label probabilities in its metadata, but no explanation.
  • DecisionModel goes wherever an evaluator takes llm, including the built-in classification evaluators.
  • OpenAI Decisions API with provider="openai", which requires openai 3.26.0 or newer.
  • TypeSafe Jev with provider="typesafe", which requires typesafe-sdk.
  • Anthropic judges use native structured output when you run language-model evaluators through the Anthropic adapter.

Configuring the Decision Model

Providers, client options, and requirements

Decision Models as a Judge

When to use a decision model instead of an LLM

Model Column and Prompt Version Timeline

November 12, 2026 Available in arize-phoenix 21.1.0+
  • See the model on the spans table. A new model column shows each span’s llm.model_name. Click a model to add an llm.model_name == "..." condition to the span filter.
  • Browse prompt versions as a timeline. The prompt overview and the Versions tab show each version with its description, author, and tags, and link to it.
  • Open a new dataset right away. Phoenix opens a dataset after you create it, and the annotation config table shows each config’s optimization direction.

px: Version Lists and Prompt Tags

November 5 to 17, 2026 Available in @arizeai/phoenix-cli 1.20.0+ (prompt-versions list), 1.21.0+ (dataset-versions list), and 1.23.0+ (prompt tag)
  • px prompt-versions list and px dataset-versions list list versions newest first, with pretty, JSON, and raw output, so you can find IDs for --version.
  • px prompt tag creates a tag on a prompt version or moves an existing tag to it. Omitting --description clears the tag’s description.

Phoenix CLI

Commands and output formats

Harbor: Model Parameters and Agent Settings

November 5, 2026 Available in arize-phoenix-client 3.7.0+ Each Harbor experiment’s harbor_agent metadata now records the model parameters, such as temperature and reasoning_effort, and the agent settings, so you can tell apart experiments that differ only in one setting. Secret-named keys and URL credentials are removed before Phoenix stores them.

Harbor

Experiment metadata fields

Additional Improvements

November 4 to 18, 2026 Available in arize-phoenix 21.1.0–21.2.0, @arizeai/phoenix-client 7.18.1–7.18.2, and arize-phoenix-otel 0.17.3
  • LLM evaluators run on newer Claude models that reject forced tool use, and on Anthropic prompts with extended thinking.
  • Retention sweeps delete traces in batches that skip locked rows, so they no longer block ingestion.
  • The standalone @arizeai/phoenix-mcp stdio package is retired in favor of the MCP server built into Phoenix at /mcp.
  • GenAI spans keep reasoning and cache-write token counts, and dataset uploads accept CSV and JSONL files with a byte order mark or blank lines.
  • The TypeScript client keeps literal $ sequences in prompt variables and keeps system instructions when converting prompts to Anthropic.
  • The Python OTel package keeps IPv6 brackets in the inferred gRPC collector endpoint.