LoginSign Up
← Back to all sparks
A

Arize Phoenix

AI-ML-PLATFORMS
Velocity5.0

AI observability and evaluation platform for LLM applications

phoenix.arize.com ↗

Phoenix 21.1 lists prompt and dataset versions from the CLI and fixes cost tracking.

llm observabilityopentelemetryclicost trackingevals
◆Current state
Phoenix moved its TypeScript tracing to OpenTelemetry JS 3 on November 11, a breaking set across phoenix-otel 3.0 and phoenix-client 8.0. The server followed on November 13 with 21.1.0: CLI commands to list prompt and dataset versions, the model shown in the spans table, stricter cost-tracking rules and a long list of server error fixes. Evals 3.10 added OpenAI Decisions and Anthropic structured output the week before.
◆Where it's heading
Phoenix is making its history scriptable (versions of prompts and datasets from the CLI) and its cost numbers trustworthy. The TypeScript SDKs are catching up to current OpenTelemetry while the server collects fixes.
◆Prediction
Expect CLI commands for more of the app's objects and follow-up TypeScript patches as users move to OpenTelemetry JS 3.

◆Recent moves

  1. 1d ago

    Phoenix 21.1: CLI version lists, model column, cost fixes

    The server release adds CLI commands to list prompt and dataset versions, shows the model in the spans table and rejects bad token rates in cost tracking. It rolls up the CLI work from 1.21 and a batch of server error fixes.

    View source ↗
  2. 3d ago

    Phoenix CLI 1.22 picks up the OpenTelemetry JS 3 client

    The CLI only moves to the new 8.0.0 client. The real change is in the otel and client packages.

    View source ↗
  3. 3d ago

    Phoenix TS client 8.0 requires OpenTelemetry JS 3 providers

    A major version because evaluateExperiment() now needs an OpenTelemetry JS 3 TracerProvider and Node.js 22.15 or newer. Users get current OpenTelemetry at the cost of a migration.

    View source ↗
  4. 3d ago

    phoenix-otel 3.0 moves to OpenTelemetry JS 3, with breaking changes

    Apps on OpenTelemetry JS 3 now have their spans recorded, but register() returns a different provider, span processors change constructors, and OTEL_* sampler and batch env vars are ignored. This is the core of the November 11 set; MIGRATION.md covers the steps.

    View source ↗
  5. 4d ago

    Phoenix CLI lists dataset versions

    px dataset-versions list shows a dataset's history with pagination and JSON output, a week after the same command for prompt versions. The CLI is steadily covering Phoenix's versioned objects.

    View source ↗
  6. 4d ago

    Evals 3.10 adds OpenAI Decisions and Anthropic structured output

    Python classification evaluators can now use OpenAI's Decisions API and Anthropic's native structured output instead of forced tool calls. Evaluators move toward provider-native ways of returning a label.

    View source ↗