Release Notes

10.16.2026: Claude Managed Agents Tracing and Gemini 4 Argon

• Claude Managed Agents tracing from a Claude Console destination and the Claude SDK
• Atomic Agents tracing with the framework’s built-in OpenInference instrumentor
• Gemini 4 Argon and Meta Muse Wonder 1.0 in the playground, with built-in cost tracking
• Server log export over OTLP, correlated with server traces
• Fixes: OAuth2 sign-in during authorization flows, Bedrock tool results, and label menus

10.13.2026: Docker Sandboxes and MCP GraphQL Tools

• Docker Sandboxes run code evaluators in Docker’s hosted sandbox environment, with optional outbound network blocking
• Claude Haiku 5.5 and Mistral pricing keep the playground and trace cost view current
• MCP GraphQL tools let agents inspect the schema, run read-only queries, and run permitted mutations on write-enabled deployments
• Analytics SQL over evaluators, prompts, trace annotations, experiment jobs, and more tables
• Evaluator fixes: metadata stays attached, and stale input mappings no longer fail runs

10.01.2026: Trace Tree Evals and Decision Spans

• Trace tree evals: annotation badges under each row, unfavorable first, with mean scores and label consensus
• DECISION spans: a dedicated icon, a detail view that names the decision model, and span-kind filters
• Project annotation configs in the TypeScript client: assign, unassign, list, and replace configs from setup scripts
• Experiment token detail charts for prompt and completion token types
• Air-gapped deployments skip the WebAssembly sandbox download and the UI’s update check

09.29.2026: Editable Dataset Examples, Trace Metrics, and Custom Skills

• Editable dataset examples — edit JSON cells in place, add and remove rows, and save the batch as one new version
• Trace tree metrics — latency, tokens, and cost under every row, with a hover preview breaking tokens and cost down by token type
• Mount your own skills with PHOENIX_SKILLS_PATHS, plus GraphQL schema search and keyboard-driven slash commands in PXI
• Decision-only evaluation models such as TypeSafe’s Jev in every TypeScript classification evaluator
• GPT-6 Sol, GPT-6 Luna, and Claude Opus 5.5 in the playground and the cost manifest
• TypeScript client and REST — secrets, dataset split writes, prompt version tags, filter expressions, and sort=start_time on the spans endpoint

09.14.2026: Coding Agent Plugins and Error Analysis

• Phoenix plugins for Claude Code, Codex, and Cursor — one install registers the MCP server and, in Claude Code and Cursor, the skills
• Error analysis served by Phoenix — the /mcp endpoint loads the skill, and PXI records its open-coding notes on the entity
• Completeness evaluator in both evals SDKs, with document relevance deprecated in favor of retrieval relevance
• Span cost filters — total_cost and cost_details, plus .identifier on annotation lookups
• Prompt version metadata over REST, GraphQL, the TypeScript client, and the playground save dialog
• REST — filter expressions for traces and sessions, and GET /v1/datasets/{dataset_identifier}/splits

09.08.2026: Harbor Plugin for Agent Benchmarks

• A job becomes experiments, one per resolved agent and model configuration, with a run per trial
• Verifier rewards become evaluations, with infra_ok separating a failed agent from a broken sandbox
• Trajectories become traces with trace_mode="atif", linked to the run, one session per trial
• Experiment names are templated from {job.name}, {agent.name}, {agent.model}, and {agent.short_digest}
• Resuming is idempotent, so a re-run tops up the same experiment

09.08.2026: New Model Providers, Trace Filters, and PXI Updates

• Z.ai, MiniMax, and Meta built in, plus gemini-3.8-flash and gpt-6-astra
• Error and latency trace filters — error, min_latency_ms, and max_latency_ms in both SDKs
• PXI — assistant settings in five linkable tabs, and slash command suggestions in the chat input
• MCP server on FastMCP 4 and the stateless protocol
• Delete traces from a project over REST within a required time window, keeping the project

09.01.2026: GitHub Issues from PXI, PII Detection, and Client Updates

• GitHub issues from PXI — duplicate search first, approval on every write, and your own token that Phoenix never stores
• PII Detection evaluator — screens conversation records including tool calls and retrieved content, in Python and TypeScript
• Prompt versions over REST — POST /v1/prompts/{prompt_identifier}/versions adds and tags a version in one call
• TypeScript client — deletePrompt, transferTraces, setProjectRetentionPolicy, getCurrentUser, and session token counts
• PXI browser scripts — one approval covers the whole script instead of a card per step
• Playground — Claude Fable 5.1 through Anthropic and AWS Bedrock

08.25.2026: Smarter PXI Workflows and Retrieval Evaluation

• PXI workflows — combine several UI actions in one turn and approve changes before Phoenix applies them
• Retrieval Relevance — score results from vector search, tools, MCP, web search, or SQL in Python and TypeScript
• Faster trace analysis — write trace filters in plain English, click a chart bin to set the time range, and compare annotations in dedicated columns
• REST administration — assign project retention policies and discover built-in or custom model providers

08.17.2026: Trace Filter Expressions and Analytics SQL

• Trace filter expressions — trace intrinsics, span rollups, root-span reach-through, and comprehensions over spans, annotations, and cost details, with span topology via children, parent_span, and siblings
• Read-only analytics SQL over MCP — describeSqlSchema publishes the queryable schema and live indexes, executeSql runs one bounded read-only statement
• Annotation details on hover — every annotation under a name, with score, label, explanation, author, and inline filter chips
• PXI approval decisions on spans — pxi.approval.decision and pxi.approval.source make accepted and rejected tool calls filterable

08.12.2026: Session Filter Expressions, REST API Expansion, and Endpoint Configuration

• Session filter expressions — session intrinsics, per-session aggregates, and comprehensions over spans, traces, annotations, and cost details, plus plain-English AI query
• Dataset splits and experiment tags over REST — create, edit, and delete splits; assign, list, and remove dataset-scoped experiment tags including baseline
• POST /v1/traces/transfer — re-parent traces into another project by trace ID or OTel trace ID
• Prompt description and metadata updates from the REST API and both client SDKs
• PHOENIX_ENDPOINT is the canonical API base URL across every SDK, the px CLI, and the MCP server
• OAuth2 workload identity — authenticate to an IDP with a platform-minted JWT instead of a client secret
• Breaking change — arize-phoenix-client 3.0.0 Google prompt helpers target google-genai

08.11.2026: Persistent Agent Sessions

• Browse and restore past chats, rename them, or mark one temporary so it is never saved
• Rewind or branch from any message — edit and re-send in place, or fork a new chat and leave the original alone
• Terminal session commands — /new, /temporary, /sessions, /model, and /compact
• Retention controls under Settings → Assistant for idle-chat deletion and a per-user cap

08.04.2026: Chat Completions Proxy, AI Query, and Annotation Charts

• OpenAI-compatible /v1/chat/completions — proxy chat completions (streaming included) to built-in or custom providers with credentials resolved on the server
• AI Query for filter fields — describe the spans, traces, or experiment runs you want in plain English and press Enter, using on-device Browser AI or any provider Phoenix knows
• Annotation metric charts — a chart per annotation name on the project Metrics page and in the chart strip, with the three-chart selection cap lifted and charts loaded on scroll
• Conversation-grounded Hallucination evaluator — judges a response against the whole transcript, including tool results; labels change to grounded / hallucinated
• Annotations in span downloads — export evaluations alongside spans as OpenInference attributes, streamed straight to disk
• Pinned note bar in span details — press n to take notes without leaving the trace

07.28.2026: Experiment Charts, Span Downloads, and Root-Span Filters

• Experiment metric charts — pin annotation score, latency, cost, token, and error-rate charts above the experiments table, drag to reorder, and read token usage broken down by sub-type
• Download spans and traces — export a selection as OTLP JSON or JSONL, or grab a single span from its detail header
• Span-ID and root-span filtering — a span_id filter across the REST API, TypeScript client, and px span list, plus a parent_span is None predicate in the span filter DSL
• Span detail as searchable tables — attributes, annotations, and notes as tables, with collapse-all and row clip/wrap controls
• Toxicity evaluator in both evals SDKs, the Monty local Python sandbox, and Claude Opus 5 in the Playground
• Breaking change — @arizeai/phoenix-client 7.0.0 and @arizeai/phoenix-evals 2.0.0 require Vercel AI SDK v7

07.22.2026: MCP Client Setup, Provider Filtering, and User Friction Evals

• One-command MCP setup — px setup mcp registers the Phoenix MCP server with Claude Code, Codex, Cursor, Gemini, OpenCode, or VS Code, plus a new Settings MCP tab with per-client instructions
• Provisioned-provider filtering — the playground model picker shows only providers that are ready to use
• User Friction evaluator — a built-in friction / no_friction classifier in both the Python and TypeScript evals SDKs
• AI SDK v7 tracing — @arizeai/phoenix-otel upgrades to OpenInference Vercel v3 (breaking; stay on 1.x for AI SDK v6)

07.17.2026: OAuth2 Authorization Server & Remote MCP

• OAuth2 authorization server — authorization-code + PKCE, dynamic client registration, refresh-token rotation with replay detection, a redesigned consent screen, and admin grant management
• Browser-based CLI login — px auth login, status, and logout obtain and silently refresh tokens per profile
• Remote MCP server (beta) — a built-in /mcp endpoint that MCP clients (Claude Code, Cursor) connect to over OAuth
• API key management — REST CRUD for user and system keys, dedicated settings tabs, and scope/audience columns
• Breaking change — GraphQL createUserApiKey/createSystemApiKey no longer accept API-key-authenticated callers

07.14.2026: Command Palette, Session Tools, and Table Customization

• Global search command palette (⌘K) — find projects, datasets, experiments, and prompts from anywhere
• Customizable data tables — pick and drag-to-reorder columns on the traces, sessions, experiments, and prompts tables
• Session tools — session stats panel, turn dividers, session-level annotation editing, and project-card session counts
• .env.phoenix credential discovery and a unified PHOENIX_PROJECT variable across the SDKs
• GPT 5.6 family in the Playground, plus a breaking change: session time-range filters now match on interval overlap (arize-phoenix 18.0.0)

07.07.2026: Metric Charts, Trace Search, and REST API Expansion

• Metric charts above tables — a resizable, per-table chart strip over spans, traces, and sessions
• Trace tree search — filter the trace tree to jump to the span you need
• Dataset labels & annotation-config assignment over REST — manage both from the API
• Classification-metric evaluators in TypeScript — precision, recall, and F-score code evaluators
• px annotation-config CLI commands and Claude Sonnet 5 in the Playground

06.30.2026: PXI Terminal Client and Annotation Summary

• PXI terminal client — launch an interactive PXI chat with npx -y @arizeai/phoenix-cli pxi; it connects to your Phoenix instance and runs a model preflight on startup, with /clear, /exit, and /help slash commands in 1.6.1+
• Annotation summary — project settings list every annotation name by count for spans, traces, and sessions, with bulk delete by name
• Copyable trace IDs — the experiment trace details dialog adds a copy-to-clipboard trace ID badge

06.26.2026: Evals as Tests — pytest and Vitest/Jest Integrations

• Familiar DX — mark tests with @pytest.mark.phoenix or import describe/test from @arizeai/phoenix-client/vitest; keep fixtures, parametrization, watch mode, and .only/.skip
• Debug in Phoenix — every case is traced; LLM-as-judge calls record under their own evaluator span
• Metrics beyond pass/fail — log scores, labels, and explanations and track them across experiments
• CI gates — per-case asserts plus suite-level acceptance criteria on aggregate scores
• Reuse evaluators — pre-built arize-phoenix-evals evaluators plug straight into a test case

06.24.2026: Annotations, Labels, and PXI Server Tools

• Trace-level annotations — trace annotation summaries appear in the trace header and project stats panel alongside span annotations
• Label management from lists — filter by label, manage labels inline, and see usage counts on the Prompts and Datasets pages, plus a prompt model column
• OAuth2 role overrides — set ROLE_RESYNC=false per IDP to keep manually assigned roles from being overwritten on login
• PXI server bash tool — subagents gain a sandboxed bash with phoenix-gql, gated network access, and a PHOENIX_AGENTS_DISABLE_BASH kill switch

06.16.2026: Time Range, Metrics, and Sharing Upgrades

• Calendar time range picker — choose an exact window from a two-month calendar with pan, zoom, and live streaming controls
• Shareable trace URLs — trace and session links carry the active time range so a copied link reproduces your view
• New project metrics — trace and session annotation scores plus prompt and completion token detail charts over time
• PXI subagents opt-in — a settings toggle controls the parallel data-retrieval helpers, off by default

06.11.2026: Time Range Selector — Search and Free-Form Durations

• Search presets — filter the preset list as you type
• Free-form durations — 25m, 2h, 3d, or last 2 hours create custom windows on the fly
• Inline editing — type start and end dates directly in the navbar selector

06.11.2026: PXI Slash Commands and Dataset Evaluator Editing

• /clear — reset the conversation into a fresh session from the chat input
• Dataset evaluators — PXI reads evaluator definitions, selects which run on the next playground experiment, and edits existing evaluators as accept/reject diffs

06.10.2026: PXI Skills Menu, Subagents, and Playground Orchestration

• Skills menu — type / to invoke skills like /debug-trace and /llm-evaluator-authoring
• Subagents — parallel, read-only data retrieval that keeps big lookups out of the main context
• Playground orchestration — load datasets, switch models, set repetitions, manage comparison instances, toggle experiment recording, and cancel runs
• Evaluator authoring and dataset management — draft LLM-judge and code evaluators, create datasets, and import spans as examples, all as reviewable diffs

06.10.2026: Claude Fable 5 in the Playground

06.02.2026: Introducing PXI

• Operates on your observability data — inspects traces, iterates on prompts as reviewable diffs, runs experiments, authors evaluators, and annotates spans
• Controls built in — state-changing actions require approval, everything runs on your deployment, and PXI can be disabled entirely

05.27.2026: Drag-to-Zoom on Project Metric Charts

• All charts covered — latency, error rate, token usage, and the trace-count sparkline all support brush selection
• Adaptive tick density — x-axis labels scale with rendered pixel width, staying readable at any zoom level

05.21.2026: Code Evaluators

• Local or hosted sandboxes — WebAssembly and Deno ship with Phoenix; E2B, Daytona, Vercel, and Modal add network and package support
• Versioned and traced — every save creates a new version, and each execution appears as a span
• Dry-run before saving — test the evaluator against a real dataset example from the UI

05.15.2026: OTel GenAI Semantic Convention Auto-Conversion

• Zero-config — point any OTel-native instrumentation at Phoenix’s OTLP endpoint; conversion happens server-side
• OpenInference wins — spans that already carry OpenInference attributes are untouched; conversion only fills in what’s missing
• Full coverage — input/output messages, tool definitions and results, system instructions, retrieval documents, usage tokens, provider, and model name

05.15.2026: Session Trace Feedback and ATIF v1.7

• Session feedback toolbar — thumbs-up/down and annotate buttons appear on each turn in Session Details; clicking creates a user_feedback trace annotation with toggle behavior
• ATIF v1.7 trajectory upload — upload_atif_trajectories_as_spans now supports embedded subagent trajectories via subagent_trajectories, trajectory_id-based deterministic span IDs for idempotent re-uploads, session_id in trajectory headers, and deterministic dispatch steps (llm_call_count: 0)

05.13.2026: Playground Thinking Controls

• Anthropic — Thinking mode (adaptive/enabled/off), token budget (min 1,024), thinking display toggle, and output effort level
• Google — Thinking budget and thinking level (LOW / MEDIUM / HIGH)

05.13.2026: Session Enhancements and Annotation Identifiers

• Expandable session turns — long messages in Session Details are now collapsed by default with an expand/collapse toggle
• Note identifier field — POST /v1/{trace,span,session}_notes accepts an optional identifier for upsert semantics; repeated calls with the same identifier overwrite the existing note
• New CLI annotation commands — px {trace,span,session}-annotations delete for bulk-remove by identifier or time range, px project get <name> to fetch project metadata; --identifier flag on annotate and add-note
• Security — f-string template formatter now blocks dunder/private attribute traversal, preventing format-string injection through user-supplied variables

05.10.2026: Playground Preferences

• Default provider and model — save a personal Playground default on the AI Providers settings page; Phoenix stores it in your browser and applies it to every new session
• Metrics aside always on — the latency/token/error metrics panel on the right side of the spans table is now visible by default, no longer gated behind a feature flag

05.08.2026: OTLP Project Routing via HTTP Header

• Works with any OTLP sender — SDK header option, Collector headers config, or OTEL_EXPORTER_OTLP_HEADERS env var
• Bug fix — deleting a dataset evaluator link no longer removes the shared evaluator row (arize-phoenix 15.5.1+)

05.05.2026: Provider Tools in Playground and Prompts

• Generic passthrough — any tool the provider SDK accepts works without a library update; verified examples include Anthropic web_search / code_execution / computer_use, OpenAI Responses web_search / file_search / code_interpreter / computer_use_preview, and Gemini google_search grounding
• Mix with function tools — provider tools and function tools coexist on the same prompt version
• SDK export — Python and TypeScript clients preserve provider tools when formatting prompts

05.05.2026: REST API Updates — Annotation Delete, Token Counts, and More

• Filter-based annotation DELETE — DELETE /v1/projects/{project}/span_annotations, trace_annotations, and session_annotations bulk-remove annotations by identifier, name, annotator_kind, or time range
• Token counts in REST — trace and session list endpoints now return cumulative_token_count_prompt, _completion, and _total fields
• Experiment CSV metadata — downloading an experiment as CSV now includes per-example dataset metadata as metadata_<key> columns
• Evals runtime capability detection — OpenAI reasoning models (o1, o3, o3-mini, o4-mini) now work with ClassificationEvaluator automatically

05.01.2026: TanStack AI Tracing

• Install — npm install --save @arizeai/openinference-tanstack-ai @tanstack/ai
• Usage — drop openInferenceMiddleware() into the middleware option of any chat() call
• Feedback welcome — this integration is brand new; please reach out via the OpenInference repo

04.30.2026: CLI Named Auth Profiles

• Commands — px profile create, list, show, edit, use, delete
• Profile status — px auth status surfaces the active profile name
• Editor autocomplete — JSON Schema published for ~/.px/profiles.json

04.30.2026: Annotation Enhancements — Sessions, TypeScript, and Identifier Queries

• CLI session annotations — px session annotate <id> and px session add-note <id> with --include-annotations / --include-notes flags on list and get
• TypeScript trace annotations — addTraceAnnotation and logTraceAnnotations exported from @arizeai/phoenix-client/traces
• addSessionNote in @arizeai/phoenix-client/sessions
• Query by identifier — GET annotation endpoints for spans, traces, and sessions accept ?identifier= to retrieve all annotations sharing a tag, project-wide

04.29.2026: Dataset Upsert

04.28.2026: Session Notes API

04.24.2026: Trace Notes API

• TypeScript — addTraceNote({ traceNote: { traceId, note } }) from @arizeai/phoenix-client/traces
• CLI — px trace add-note <trace-id> --text "..." and --include-notes flag on px trace get / px trace list
• Reserved name — note is no longer accepted on the generic annotation endpoints; use POST /v1/trace_notes

04.24.2026: arize-phoenix-otel 0.16

• Context managers — using_session, using_user, using_metadata, using_tags, using_attributes, using_prompt_template, suppress_tracing (usable as with blocks or decorators)
• Semantic conventions — SpanAttributes, OpenInferenceSpanKindValues, OpenInferenceMimeTypeValues
• Single install — pip install "arize-phoenix-otel>=0.16.0" is enough for register() + context propagation

04.22.2026: Secrets Settings Page

• Add / Replace / Delete secrets from the browser
• Search and filter the secrets list by key name or owner
• Admin-only — all mutations require admin access

04.22.2026: `trace_id` in Experiment Evaluators

• Optional — evaluators without trace_id are unaffected
• Works sync and async — supported on both function-based and protocol-based evaluators

04.20.2026: Span Attribute Filtering

• Python — client.spans.get_spans(..., attributes={"llm.model_name": "gpt-4o"})
• TypeScript — getSpans({ ..., attributes: { "llm.model_name": "gpt-4o" } })
• CLI — px span list --attribute "llm.model_name:gpt-4o"
• Type-aware — int, float, bool, and str values each match their corresponding stored type

04.20.2026: CLI Span Notes

04.20.2026: Claude Opus 4.7 in the Playground

04.16.2026: Azure Managed Identity for PostgreSQL

• Zero-credential setup — tokens are fetched and refreshed automatically on each connection
• New [azure] extra — pip install 'arize-phoenix[azure]' pulls in azure-identity and aiohttp
• PHOENIX_POSTGRES_AWS_IAM_TOKEN_LIFETIME_SECONDS deprecated — silently ignored with a startup warning

04.14.2026: CLI Annotation Commands

• px span annotate <span-id> — attach a label, score, or explanation to any span by OTel span ID
• px trace annotate <trace-id> — annotate a full trace by OTel trace ID
• Annotator kinds — HUMAN, LLM, or CODE; submitting again with the same name updates the existing entry

04.13.2026: @arizeai/phoenix-otel 1.0

• Tracing helpers — withSpan, traceChain, traceAgent, traceTool wrap functions with OpenInference spans and follow global provider changes
• observe decorator — trace class methods with TypeScript 5+ standard decorators while preserving this
• Context setters — setSession, setUser, setMetadata, setTags, setAttributes, setPromptTemplate propagate attributes to child spans
• Attribute builders — getLLMAttributes, getRetrieverAttributes, getEmbeddingAttributes, getToolAttributes, and more for raw OTel spans
• Semantic conventions — SemanticConventions and OpenInferenceSpanKind re-exported; no second dependency needed
• Redaction — OITracer + traceConfig or OPENINFERENCE_HIDE_* env vars strip sensitive attributes before export

04.10.2026: Shareable Project URLs

• Project by name — /redirects/projects/default resolves and redirects to the project page
• Other patterns — traces, spans, sessions, and prompt tags are also supported via /redirects/...

04.07.2026: PostgreSQL Read Replica Routing

• Set and go — point the env var at your replica; writes always go to the primary
• No config change needed when a replica is not configured — falls back to primary

04.07.2026: Breaking Changes — Phoenix v14

• CLI is now subcommand-first — phoenix serve --dev replaces phoenix --dev serve
• px.Client() removed — use from phoenix.client import Client with base_url= instead of endpoint=
• /v1/evaluations removed — use /v1/span_annotations, /v1/trace_annotations, or /v1/document_annotations
• Evals 1.0 removed — arize-phoenix-evals 3.0.0 drops the legacy/ subpackage; phoenix.experiments is replaced by phoenix.client.experiments
• GraphQL pagination requires first — Project.spans, Trace.spans, and ProjectSession.traces now require an explicit first argument (max 1000)
• Resume interrupted SDK experiments — continue missing or failed runs with resume_experiment and async_resume_experiment

04.03.2026: ATIF Trajectory Upload

• Supports ATIF v1.0–v1.6 including multimodal content (images) in v1.6+
• Subagent linking — upload parent and child trajectories together to link them into one trace
• Idempotent — trace/span IDs are derived from session_id via SHA-256; re-uploading the same file is safe

04.01.2026: Python 3.14 Support

04.01.2026: Secrets Management REST API

• PUT /v1/secrets — batch upsert/delete with value: null to remove a key
• Admin-only, atomic, silent no-op for deleting non-existent keys

04.01.2026: `get_traces` — Retrieve Traces via Python SDK

• include_spans=True to embed full span detail in each trace
• session_id to filter to one or more sessions
• Async variant available on AsyncClient

03.30.2026: Delete Prompts and Prompt Version Tags

• DELETE /v1/prompts/{prompt_identifier} — permanently deletes a prompt and all its versions, tags, and labels
• DELETE /v1/prompt_versions/{id}/tags/{tag_name} — removes a single tag from a specific prompt version

03.24.2026: Prompt Version Diff View

• Side-by-side diff across the full chat template, including all content part types

03.24.2026: Evals Accept Structured Data Inputs

• Structured inputs (dicts, lists) are JSON-serialized before prompt rendering — no manual json.dumps() needed
• LLM invocation kwargs accepted by all built-in evaluators (FaithfulnessEvaluator, CorrectnessEvaluator, etc.)

03.23.2026: `px auth status` Shows Username and Role

03.22.2026: `px spans` — Fetch and Filter Spans from the CLI

• --span-kind, --status-code, --name, --trace-id filters
• --last-n-minutes / --since time range controls
• --include-annotations to attach span annotations to the output

03.22.2026: `px self update` and `GET /v1/user`

03.18.2026: Span Filters — Name, Kind, and Status Code

• name filter to match one or more span names
• span_kind filter for LLM, CHAIN, TOOL, RETRIEVER, and other span kinds
• status_code filter to isolate error, success, or unset spans
• Available in Python via client.spans.get_spans(name=..., span_kind=..., status_code=...)
• Available in TypeScript via getSpans({ name, spanKind, statusCode })

03.13.2026: List Traces by Project

• GET /v1/projects/{project}/traces with sort, order, limit, cursor, include_spans, and session_identifier params

03.11.2026: Session Turns API

• Chronological turn ordering from session traces sorted by start time
• SessionTurnIO with MIME type — supports text/plain, application/json, and image types
• Batched root span fetching with pagination to handle large sessions
• Async variants available in both Python and TypeScript clients

03.10.2026: Session Management REST APIs

• Single session retrieval by ID or GlobalID with optional project filtering
• Bulk session listing with pagination, project filtering, and sorting
• Session deletion with automatic cascade through traces and spans
• DataFrame export in Python for sessions data analysis
• Configurable timeouts for all session operations

03.09.2026: Filter Spans by Trace ID and Parent Relationships

• Trace ID filtering to retrieve spans from specific traces
• Parent ID filtering to query root spans or span children
• Multi-trace queries via repeated trace ID parameters
• Composable with existing filters like time ranges and limits

03.04.2026: Incremental Evaluation Metrics in Playground

03.04.2026: Brute Force Login Protection

03.07.2026: Unified Dataset Upload with Drag-and-Drop

• Automatic format detection for CSV and JSONL files
• Drag-and-drop file selection with visual feedback
• Streaming parser that handles large files efficiently
• RFC 4180 CSV support including quoted fields, escaped quotes, and BOM handling
• Detailed error messages for parsing issues with line-by-line feedback

03.10.2026: Drag-and-Drop Column Assignment for Datasets

• Visual column assignment with draggable chips and drop targets
• Smart auto-assignment based on column names like “input”, “output”, “reference”
• Live dataset preview showing the final structure as you make changes
• Keyboard navigation support for accessibility
• Raw data preview in tabular format alongside final dataset view

03.08.2026: Extended Model Provider Support

• Perplexity AI for research and web-grounded responses
• Together AI with models from Moonshot, DeepSeek, Qwen, and GLM
• Cerebras for ultra-fast inference with Llama models
• Fireworks AI with Llama 4 Scout and Maverick variants
• Groq for low-latency Llama and Qwen deployments
• Moonshot (Kimi) with extended 128k and 32k context models
• Cost tracking enabled for Cerebras, Fireworks, Groq, and Moonshot

03.10.2026: Provider Visibility Controls

• Allow-list mode to show only specified providers
• Case-insensitive configuration with typo detection warnings
• Set to NONE to hide all providers from the UI

03.08.2026: Latest OpenAI GPT Models

03.08.2026: Project Editing from Settings

03.11.2026: Breaking Change: Removed Deprecated Annotations API

03.05.2026: SDK Session Retrieval

02.27.2026: Sessions API and CLI Support

02.24.2026: Claude Agent SDK Integration

• Install @arizeai/openinference-instrumentation-claude-agent-sdk alongside @arizeai/phoenix-otel
• See agent execution flows and tool invocations in Phoenix’s trace UI
• Get started →

02.14.2026: Phoenix 13.0

• Attach evaluator suites directly to datasets and run them server-side on every Playground experiment.
• Reuse server-managed custom providers (OpenAI, Azure OpenAI, Anthropic, AWS Bedrock, Google GenAI) across Playground, prompts, and dataset evaluators.
• Choose OpenAI API type per configuration (chat.completions.create or responses.create) with automatic parameter compatibility handling.
• Use new Playground workflows such as cancellation, template variable autocomplete, appended messages, improved prompt selection, and URL state for prompt IDs/versions/tags.
• Get expanded dataset and experiment ergonomics, model/provider updates (including Claude Opus 4.6 and Azure OpenAI v1 migration), and infrastructure improvements like a session ID index for spans.

02.12.2026: OpenAI Responses API Type Support

• API type selection: Choose Chat Completions or Responses per model configuration.
• Custom provider support: OpenAI and Azure OpenAI custom providers can be configured with an API type for consistent routing.
• Parameter compatibility: Phoenix maps shared invocation parameters to the chosen API type and filters unsupported fields automatically.

02.12.2026: Dataset Evaluators

• Attach once, evaluate everywhere: Add LLM or built-in code evaluators to a dataset and reuse them across Playground experiments.
• Flexible input mapping: Map evaluator inputs to dataset fields so each example is evaluated consistently.
• Built-in visibility: Each evaluator captures traces for debugging and refinement, with details available from the evaluator view.

02.11.2026: Custom Providers for Playground and Prompts

• Centralized configuration: Manage provider credentials and routing in Settings and reuse them across the playground and prompt versions.
• SDK-specific authentication: Support API keys, Azure AD token providers, or default credentials (Azure/AWS) depending on the SDK.
• Model selection integration: Custom providers show up in model menus as their own provider group and inherit model listings from the underlying SDK.
• Request-level overrides: Continue to supply custom request headers per prompt while using custom provider configuration for routing and authentication.

02.09.2026: Claude Opus 4.6 Model Support

• Anthropic provider integration: Access Claude Opus 4.6 directly through the playground with the thinking invocation parameter enabled for extended reasoning workflows
• AWS Bedrock support: Deploy Opus 4.6 through Bedrock with the region-specific model identifier
• Automatic cost tracking: Token costs are calculated using the latest pricing ($5 per million input tokens, $25 per million output tokens, plus cache read/write rates)

01.31.2026: Tool Selection and Tool Invocation Evaluators

• Identify tool selection errors where agents choose suboptimal or incorrect tools
• Debug parameter issues including hallucinated fields, malformed JSON, and incorrect values
• Improve tool descriptions and agent prompts based on systematic evaluation
• Validate multi-tool and multi-turn interactions across complex agent workflows

01.28.2026: Configurable Email Extraction for OAuth2 Providers

</ReleaseUpdate>

<ReleaseUpdate label="02.01.2026" href="/docs/phoenix/release-notes/02-2026/02-01-2026-tool-selection-and-tool-invocation-evaluators" title="Extract from Azure AD preferred_username claim">
PHOENIX_OAUTH2_AZURE_AD_EMAIL_ATTRIBUTE_PATH=preferred_username
</ReleaseUpdate>

<ReleaseUpdate label="02.01.2026" href="/docs/phoenix/release-notes/02-2026/02-01-2026-tool-selection-and-tool-invocation-evaluators" title="Extract from nested claims">
PHOENIX_OAUTH2_CUSTOM_IDP_EMAIL_ATTRIBUTE_PATH=user.contact.email

01.22.2026: CLI Commands for Prompts, Datasets, and Experiments

• List and view prompts with px prompts and px prompt <name>
• Pipe prompts to AI assistants for optimization and analysis
• Text format output with XML-style role tags for LLM consumption

• Create and manage datasets with px datasets and px dataset <name>
• Add examples and query dataset contents
• Export datasets for offline analysis

• Run experiments and compare results across configurations
• View experiment details and performance metrics
• Track changes across prompt and model variations

01.23.2026: CLI Authentication Configuration

01.21.2026: Create Datasets from Traces with Span Associations

from phoenix.client import Client

client = Client()
dataset = client.datasets.create_dataset(
    name="production-queries",
    dataframe=spans_df,
    input_keys=["input"],
    output_keys=["output"],
    span_id_key="context.span_id"  # Links examples to spans
)

import { createClient } from '@arizeai/phoenix-client';

const client = createClient();
await client.createDataset({
    name: "production-queries",
    examples: examples.map(ex => ({
        input: ex.input,
        output: ex.output,
        spanId: ex.spanId  // Preserves trace links
    }))
});

• Batch resolution of span IDs for optimal performance
• Graceful fallback when span IDs are missing or invalid
• Backwards compatible with existing dataset creation workflows
• Bidirectional navigation between evaluation results and production traces

01.19.2026: Export Annotations with Traces

px traces export --include-annotations > traces_with_feedback.jsonl

01.22.2026: CLI Prompt Commands: Pipe Prompts to AI Assistants

01.21.2026: Create Datasets from Traces with Span Associations

01.21.2026: Phoenix CLI: Datasets, Experiments & Annotations

01.17.2026: Phoenix CLI: Terminal Access for AI Coding Assistants

01.05.2026: Appended Messages for Playground Experiments

12.20.2025: Improved User Preferences

12.12.2025: Support for Gemini Tool Calls

12.09.2025: Span Notes API

12.06.2025: LDAP Authentication Support

12.04.2025: Evaluator Message Formats

12.03.2025: TypeScript createEvaluator

12.01.2025: Splits on Experiments Table

11.29.2025: Add support for Claude Opus 4-5

11.27.2025: Show Server Credential Setup in Playground API Keys

11.25.2025: Split Assignments When Uploading a Dataset

11.23.2025: Repetitions for Manual Playground Invocations

11.14.2025: Expanded Provider Support with OpenAI 5.1 + Gemini 3

11.12.2025: Updated Anthropic Model List

11.09.2025: OpenInference TypeScript 2.0

• Added easy manual instrumentation with the same decorators, wrappers, and attribute helpers found in the Python openinference-instrumentation package.
• Introduced function tracing utilities that automatically create spans for sync/async function execution, including specialized wrappers for chains, agents, and tools.
• Added decorator-based method tracing, enabling automatic span creation on class methods via the @observe decorator.
• Expanded attribute helper utilities for standardized OpenTelemetry metadata creation, including helpers for inputs/outputs, LLM operations, embeddings, retrievers, and tool definitions.
• Overall, tracing workflows, agent behavior, and external tool calls is now significantly simpler and more consistent across languages.

11.07.2025: Timezone Preference

11.05.2025: Metadata for Prompts

11.03.2025: Playground Dataset Label Display

11.01.2025: Resume Experiments and Evaluations

10.30.2025: Metadata Support for Experiment Run Annotations

10.28.2025: Enable AWS IAM Auth for DB Configuration

10.26.2025: Add Split Edit Menu to Examples

10.24.2025: Filter Prompts Page by Label

10.20.2025: Splits

10.18.2025: Filter Annotations in Compare Experiments Slideover

10.15.2025: Enhanced Filtering for Examples Table

10.13.2025: View Traces in Compare Experiments

10.10.2025: Viewer Role

10.08.2025: Dataset Labels

10.06.2025: Paginate Compare Experiments

10.05.2025: Load Prompt by Tag into Playground

10.03.2025: Prompt Version Editing in Playground

09.29.2025: Day 0 support for Claude Sonnet 4.5

09.27.2025: Dataset Splits

09.26.2025: Session Annotations 🗂️

09.25.2025: Repetitions

09.24.2025: Custom HTTP headers for requests in Playground

09.23.2025: Repetitions in experiment compare slideover

09.22.2025: Helm configurable image registry & IPv6 support

09.17.2025: Experiment compare details slideover in list view

09.15.2025: Prompt Labels 🏷️

09.12.2025: Enable Paging in Experiment Compare Details 📄