​

Breaking Changes in the GraphQL API

• rootSpansOnly and orphanSpanAsRootSpan are removed from Project.spans and Trace.spans. Use filterCondition: "parent_id is None" instead, or "parent_span is None" to count orphans as roots.
• Project.traceAnnotationsNames is now Project.traceAnnotationNames.
• applyDatasetExampleChanges is now patchDatasetExamples. It takes an ordered, JSON Patch-style list of operations that commits as one dataset version, and requires datasetId.

​

Edit Dataset Examples in Place

• Edit JSON cells in place, add rows with an optional custom ID, and remove or restore rows before you save. Phoenix flags duplicate custom IDs before the save goes through.
• Save with ⌘S. The toolbar counts pending changes, the save dialog takes an optional version description, and Phoenix warns before you leave the page with unsaved edits.
• Search and split filters stay live while you edit, and the toolbar reports any changes the current search hides.

​

Latency, Tokens, and Cost on Every Trace Tree Row

• Hover a row to preview its breakdown. One preview follows the pointer down the tree and loads details only once the pointer settles, so large traces stay responsive.
• Tokens and cost share one breakdown. Each measure splits into token types, so you can compare a type’s share of tokens with its share of cost.

​

PXI: Custom Skills, Schema Search, and Slash-Command Keys

PHOENIX_SKILLS_PATHS=/opt/skills/team-analysis,/srv/agents/skills

• Each path is a skill or a directory of skills. A skill’s directory name must match the name in its SKILL.md frontmatter, and the startup banner lists what mounted. Use absolute paths in deployments.
• PXI searches the GraphQL schema instead of reading the full SDL: phoenix-gql schema --search <terms> returns ranked field signatures, and phoenix-gql schema --names <Type.field> prints one definition and the paths that reach it.
• Slash-command hints respond to the keyboard in the pxi terminal client. Arrow keys move the highlight, Tab completes, and Enter runs, so /he plus Enter runs /help.

​

Score Classifications with Decision-Only Models

import { typeSafeAi } from "@ai-sdk/typesafe-ai";
import { createHallucinationEvaluator } from "@arizeai/phoenix-evals";

const evaluator = createHallucinationEvaluator({
  model: typeSafeAi.evaluationModel("jev-latest"),
});

const result = await evaluator.evaluate({
  input: "Retrieved document: Arize Phoenix is open source.\n\nUser: Is Arize Phoenix open source?",
  output: "Yes, Arize Phoenix is open source.",
});
console.log(result.label, result.metadata?.probabilities);

• Same evaluators, same labels. Phoenix detects the evaluation model and sends the call through experimental_evaluate instead of generateObject.
• Results carry a label and score but no explanation. metadata reports the full label distribution as probabilities, along with the modelId.

​

GPT-6 Sol, GPT-6 Luna, and Claude Opus 5.5

• GPT-6 Sol and GPT-6 Luna are reasoning models, so Phoenix sends them to the Responses API on OpenAI and Azure OpenAI, where tool calls work.
• Claude Opus 5.5 always uses adaptive thinking, so the playground omits temperature, top-p, top-k, and extended-thinking settings. Bedrock lists it as anthropic.claude-opus-5-5.
• Opus 5.5 spans bill at Opus 5.5 rates. They previously matched the claude-opus-5 pattern and billed at Opus 5’s higher rates.

​

TypeScript Client: Secrets, Splits, Prompt Tags, and Filters

import { createDatasetSplit } from "@arizeai/phoenix-client/datasets";
import { upsertPromptVersionTag } from "@arizeai/phoenix-client/prompts";
import { upsertOrDeleteSecrets } from "@arizeai/phoenix-client/secrets";
import { getTraces } from "@arizeai/phoenix-client/traces";

const apiKey = process.env.OPENAI_API_KEY;
if (!apiKey) throw new Error("OPENAI_API_KEY is required");

await upsertOrDeleteSecrets({
  secrets: [
    { key: "OPENAI_API_KEY", value: apiKey },
    { key: "OLD_PROVIDER_API_KEY", value: null },
  ],
});

await createDatasetSplit({
  dataset: { datasetName: "support-tickets" },
  name: "regression",
  exampleIds: ["ticket-14", "ticket-22"],
});

await upsertPromptVersionTag({
  prompt: { versionId: "UHJvbXB0VmVyc2lvbjox" },
  name: "production",
});

const { traces } = await getTraces({
  project: { projectName: "my-project" },
  filter: "error_count > 0 and total_cost > 0.05",
});

• upsertOrDeleteSecrets (7.13.0) applies a batch atomically: a value creates or updates a key, and null deletes it. The response lists key names only, never values.
• createDatasetSplit, updateDatasetSplit, and deleteDatasetSplit (7.14.0) manage splits on a dataset selected by name or GlobalID. Requires Phoenix server 19.20.0+.
• upsertPromptVersionTag and deletePromptVersionTag (7.15.0) create, move, and remove prompt version tags.
• filter on getTraces and listSessions (7.12.0) takes the same expressions as the UI and deprecates the error, minLatencyMs, and maxLatencyMs parameters. Requires Phoenix server 20.12.0+.

​

REST API Updates

• GET /v1/projects/{project_identifier}/spans accepts sort=start_time, so a page limit returns the latest spans rather than the most recently ingested ones. The default sort=id keeps insertion order.

​

Additional Improvements

• Experiment names and descriptions are editable from the experiments table and when you launch an experiment from the playground.
• register() no longer crashes with opentelemetry-exporter-otlp-proto-http 1.45.
• Analytics SQL on SQLite supports the bundled time functions (time_trunc, time_parse, time_fmt_datetime, and time_sub), and latency_ms is nanosecond-exact.
• PXI streams faster, keeps partial output when you interrupt a turn, and closes the turn’s trace when a server tool raises.
• PromptTemplate message lists accept role aliases such as developer, human, ai, and model.
• The rate limiter rejects an invalid initial request rate up front.
• SQLite autoincrement counters survive migrations that rebuild a table, so upgrades never reuse a primary key.
• Metric chart time ranges stay stable, and the table’s load-more control stays in the footer in Safari.
• The cost manifest carries refreshed token prices for built-in models.