📊 Phoenix 17.7.0 makes your token usage legible. New token detail charts break prompt + completion tokens into their parts, over time: • Prompt → input, cache read, cache write, audio • Completion → output, reasoning, audio

Phoenix just hit 10,000 GitHub stars! Three years ago, Phoenix didn't exist. Arize was a closed-source company. A small team was asked to change that. Catch the full interview with the team who made it happen and where AI observability is going next: arize.com/phoenix-10k

"Don't trust. Evaluate." @nearestnabors.com set out to replace Sonnet with Gemma. The evals showed a quantifiably better option. Full walkthrough: capability evals + prompt engineering to ship a local 3B that matches Sonnet, 2x faster, $0/call. Built with Phoenix. arize.com/blog/how-to...

Phoenix now lets you compose evaluation strategies in code. Most eval tooling hands you a fixed menu of judge templates. Real evaluation is rarely that tidy.

The Arize DevRel team wants to connect with Phoenix users like you. What you're tracing, what's working, what's rough? Schedule time with the team here: cal.com/team/arize-...

Something we’ve been playing with and liking a lot: Give every coding agent its own observability stack.

A comprehensive 2-hour evaluations workshop, for free! At AI Engineer: Europe, head of DevRel Laurie Voss gave this workshop that covers: - What is an eval? - Why are they important? - How and why to manually examine the data - How to write your own evals www.youtube.com/watch?v=Xfl...

Phoenix 15.5 → 15.7 🏷️ Note identifiers for upsert semantics (15.7)

Agents are becoming increasingly distributed. A single agentic workflow might cross an app server, an MCP tool server, another agent over A2A, an ACP-compatible runtime, a vector database, an LLM provider, and multiple internal APIs.

Evals should help you ship FASTER. They are not there to gatekeep . They should be an accelerant: a way to move faster with more confidence. That only works if feedback is fast and accurate. When evals are slow, teams run them less often, test fewer cases, and push quality checks to later.

The official tanstack AI Otel support is out! Looking for a OSS backend for traces, datasets, and replay? Check out our docs for how to seamlessly add debug-ability. x.com/tan_stack/s...

One of the oldest lessons in ML is still one of the most useful for working with LLM apps: Don’t evaluate on the same data used to build. Train/dev/validation/test splits exist for a reason. They help separate “this worked because it was tuned against it” from “this actually generalizes.”

Who judges the evaluators?

Coding agents are growing ~10x year over year. But most are still operating blind. If agents are shipping to production, they need to be grounded in real infrastructure: * Traces → what actually happened * Evaluations → what’s correct vs broken * Feedback → what to fix next

A heavy shipping week for Phoenix (04.29 → 05.05). Eight items landed. The four worth pulling out for anyone running production eval pipelines: ↓ → Provider tools in Playground & Prompts

Phoenix shipping report, 04.29 → 05.05. provider tools, annotation cleanup, and a real CLI profiles system landed. quick rundown:

Different ways to deploy Arize Phoenix — from local dev to production scale. Phoenix is a lightweight, containerized platform for LLM tracing and evaluation. How you deploy it should match your team size, scale, and reliability needs. Here are the most common patterns:

Phoenix 15.4 now supports provider tools A concrete example of why provider tools matter for grounding and tool-call quality.

LLM applications are noisy by default. Non-determinism, open-ended inputs, and multi-step tool use create a flood of traces that all look “fine” — until something breaks. The difference between noise and signal is annotation.

Most tool-calling evals treat everything as one check. But a tool call actually has 3 separate steps — and each can break independently 👇 🧠 Tool Selection Did the model pick the right tool? ⚙️ Tool Parameters Did it call that tool correctly (valid args, right values)?

🚀 Making your traces smarter with Context Propagation in Phoenix (Open Source) If you're building LLM-powered apps or distributed systems, you’ve probably hit this problem: 👉 “How do I consistently track users, sessions, and metadata across everything?”

Most LLM observability breaks down because we mix levels of abstraction. OTEL gives spans and traces; Phoenix adds sessions. Together they form a hierarchy.

The biggest bottleneck in AI application development right now isn’t models — it’s evaluation. Every AI engineer faces the same question: Which prompt or model actually gives the best balance of quality, latency, and cost? Without strong evaluation, you’re guessing. ---

Your agent eval could be grading the wrong thing. Output-only evals check the final answer. But agent failures usually happen mid-trajectory — the model picked the wrong tool, skipped retrieval, looped on an LLM call. The output can still look reasonable while the path was completely broken.

TanStack just shipped TanStack AI, a typed, streaming, tool-calling chat layer for TypeScript. Today we’re releasing the OpenInference middleware for TanStack AI: drop it into the middleware option on chat() and every run streams into Phoenix as fully-shaped traces! arize.com/docs/phoeni...

Will your harness survive the next model release? Build the suite, not just the loop. @nearestnabors.com built three mini harnesses, plugged them into Phoenix, and ran 117 evals across Claude, GPT-4o, and Gemma 4. The repo is forkable so you can drop in your own loop: x.com/rachelnabor...

Your traces just got a notebook. We shipped span and trace notes in Phoenix — and the unlock isn't really for humans. It's for agents.

Three years ago we started OpenInference because there was no standard way to trace an LLM call, let alone an agent. OpenTelemetry was always the right home — same primitives as the rest of the stack, no parallel system bolted on for AI. www.youtube.com/watch?v=nLH...

OpenInference instrumentation updates are out (Apr 29): github.com/Arize-ai/op... A set of focused fixes across the Python packages, plus a breaking MCP change: python-openinference-instrumentation (0.1.48)

We recently recorded a session with Hamel Husain on what evals and observability look like when coding agents start doing the coding. youtu.be/AsDjsLvtkbQ

Shipping in Phoenix: a new Traces view for sessions.

Release notes — April 27, 2026 A few things shipped in Phoenix over the past week: - Span attribute filtering across the stack — pass attribute filters from the Python client, TypeScript client, REST API, or CLI. Type-aware (int/float/bool/str each match their stored type). Filters AND together.

Manual LLM tracing just got one install lighter. arize-phoenix-otel 0.16 re-exports OpenInference context managers and semantic conventions straight from phoenix.otel. from phoenix.otel import register, using_session, SpanAttributes arize.com/docs/phoeni...

Shipped this week: • Filter spans by attribute — Python, TS, REST, CLI • Secrets UI — manage API keys without curl • Claude Opus 4.7 in the Playground • trace_id in experiment evaluators • Azure Managed Identity for Postgres • CLI span notes for open coding arize.com/docs/phoeni...

Quis custodiet ipsos custodes? Your judge is just another LLM app. Phoenix traces evaluators so you can: → audit decisions → refine prompts → build benchmark datasets → curate fine-tuning data for smaller judges The eval loop is the AI eng loop. Treat it like one.

Get visibility into agent benchmarking execution using ATIF

Phoenix 14 introduces Experiment Jobs. Hit Run and Phoenix queues an ExperimentJob per instance. A background daemon fans them out to workers and streams progress back.

phoenix CLI and REST API now support pulling spans by OTEL attribute values so you can start debugging targeted parts of your agent topology.

When debugging agents, you need to get to the problematic issues QUICKLY! Phoenix now comes with quick list-details navigations of conversations with agents as well as resizable drawers so you can quickly navigate between conversations as well. Also sessions now supports VIM bindings for pagination!

@arizeai/phoenix-otel 1.0 is out. One import to trace TypeScript agents. Wrap a function, get a span — inputs, outputs, errors, span kind, all recorded automatically.

Some quality of life improvements Arize Phoenix now supports python 3.14 across all SDKs and has new REST endpoints including API / Secret rotation APIs. arize.com/docs/phoeni...

Guess how many 🌟 Phoenix has on GitHub? Yes, this meme has been all over our internal slack today. github.com/Arize-ai/ph...

We just added streamdown rendering for our LLM outputs now look and perform so well. Particularly loving the bash and code rendering 🤟

Post 1 PSA: LiteLLM versions 1.82.7 and 1.82.8 on PyPI were compromised with a credential-stealing payload. Are you using Phoenix's optional LiteLLM extra for phoenix-evals directly or through DSPy, Smolagents, or CrewAI? Did you install or upgrade to litellm 1.82.7 or 1.82.8 from PyPI? See: 🧵

Track exactly what changed in your prompts with diffs!

Arize Phoenix 13.11.0 — Adds @perplexity_ai and @togethercompute as built-in providers for benchmarking and evaluation.

Arize AI Phoenix v13.10 now supports Cerebras, Fireworks AI, Groq, and Moonshot (Kimi), as well as OpenAI's GPT 5.4 models, allowing you to compare hundreds of more models side by side for benchmarking, task evaluation, or LLM judge building. #AI #LLM #OpenSource #Observability #Evals

Anthropic's Claude Agent SDK lets you build AI agents that autonomously read files, run commands, search the web, and edit code — the same tools and agent loop that power Claude Code, now programmable in Python and TypeScript. Docs: arize.com/docs/phoeni...

This is the chart that should make every AI engineer pause. In the demo agent we evaluated: ⚪Tool selection: 100% ⚪Matches expected tool calls: 36%

Sessions are how your users actually experience your AI app — not as isolated calls, but as multi-turn conversations. Now available in the Phoenix SDKs. arize.com/docs/phoeni...