Skip to main content

Phoenix now shows eval results where you debug traces: on the span rows themselves. You can scan a trace for failed or mixed evals before opening a span, and decision-model calls now appear as DECISION spans instead of looking like ordinary LLM calls.

Warning

Breaking change in arize-phoenix 20.17.0: Python 3.11 or newer is now required. arize-phoenix no longer installs on Python 3.10, so plan the runtime upgrade before installing this release. SQLAlchemy 2.1.1 or newer is also required. In filters, a JSON number now matches its string spelling on both SQLite and PostgreSQL, and on SQLite 1 and 1.0 no longer compare equal.

Eval Results in the Trace Tree#

September 30, 2026

Available in arize-phoenix 20.18.0+

Large traces are hard to debug when eval results only appear after you open each span. Phoenix now puts annotation summaries under each span row, so you can start with the bad eval and then open the span that caused it.

Phoenix trace tree with evaluation badges under span rows
Eval badges appear under span rows in the trace tree, so failed and mixed evals are visible while you scan.
  • Scan before opening a span. Each row can show compact badges with the mean score and the most common label for its annotations.
  • Start with the failures. Unfavorable results appear first, and extra badges collapse behind a +N count when the tree is narrow.
  • Hover a row for detail. The span preview lists the full annotation set next to the token and cost breakdown.
  • See new eval configs right away. When you link an annotation config to a project, its results show up in the trace tree without a page reload.
Annotate Traces

Add human and LLM annotations to spans

Evaluate Phoenix Traces

Run evals over traces and log the results

Decision Spans#

October 1, 2026

Available in arize-phoenix 20.19.0+

Routers, classifiers, and rubric scorers do a different job than chat models. Phoenix now recognizes the OpenInference DECISION span kind for calls that score or pick among options, so these calls are easier to find and no longer look like normal text generation.

Phoenix trace with an OpenAI Decisions span selected, showing the decision badge and the gpt-6-luna model name
Decision spans get their own icon in the trace tree and a detail view that names the decision model.
  • Decision spans get their own icon and detail view. The detail view shows the decision input and output cards instead of the LLM message layout.
  • The span view names the decision model. Phoenix uses decision.model_name when it is set, then the provider-reported model, then the requested model.
  • Filter by DECISION anywhere you inspect spans. The span kind works in the UI, the REST API, and px span list --span-kind DECISION.
  • OpenInference instrumentors emit decision spans for the OpenAI Decisions API and TypeSafe AI System One, another decision-model provider. The decision.* attributes may still change as more decision-model providers appear.
Span Kinds

DECISION and the other OpenInference span kinds

OpenAI Tracing

Trace the OpenAI Decisions API

TypeScript Client: Project Annotation Configs#

September 30, 2026

Available in @arizeai/phoenix-client 7.16.0+ (requires Phoenix server 17.16.0+)

Teams often assign different eval configs to different projects. The TypeScript client can now manage those assignments from setup scripts or CI, so a project shows the evals that matter for that workflow as soon as traces arrive.

  • assignProjectAnnotationConfig is safe to rerun. Select the config by configName, configId, or config. Use configId when the name contains /.
  • setProjectAnnotationConfigs replaces the assignment set. It adds and removes assignments to match configIds, and it never deletes the configs themselves.
  • listProjectAnnotationConfigs returns the full set. The helper pages through every assigned config for you.
Projects API Reference

Every helper on the projects subpath

Additional Improvements#

September 30 to October 1, 2026

Available in arize-phoenix 20.17.0–20.19.0

  • Experiment metrics include prompt and completion token detail charts, which split tokens into input, cache, output, reasoning, and audio parts across recent experiments.

    Prompt token details chart showing cache read and input tokens across seven experiments
    The prompt token details chart splits each experiment's prompt tokens into input and cache reads.

  • PHOENIX_ALLOW_EXTERNAL_RESOURCES=false now skips the WebAssembly sandbox download and the UI’s GitHub star count and update check. Set PHOENIX_WASM_BINARY_PATH to use the WebAssembly sandbox offline.

  • Dataset.experiments in GraphQL accepts sort and sequenceNumbers, so you can fetch experiments by their dataset sequence number without paging.

  • PXI marks a bash tool span as an error when its command exits with a non-zero code.

  • Live streaming pauses while a trace or session drawer is open, so the tables behind it stop refreshing.

  • Experiment JSON and CSV exports no longer fail when a run errored. Errored runs export with a null output.

  • Spans that send both OpenInference and gen_ai.* messages no longer get garbled messages. Phoenix keeps the instrumentation’s own message list instead of mixing the two.

  • Evaluator input mappings keep their Path or Text mode when you erase and retype a template variable.

  • Categorical annotation configs ignore blank category rows when you save them, and saving a playground prompt to an existing prompt prefills metadata from its latest version.