arize-phoenix

As agents gain capabilities due to skills, MCPs, and codemode, evaluating agents becomes increasingly more difficult because you need to test across multiple environments and harnesses. Arize Phoenix your benchmarking Harbor runs so pinpointing improvements and regressions becomes tractable.

your agent can think fast and slow. kahneman's "thinking, fast and slow" describes two modes: system one: fast, cheap, almost automatic (jev) system two: slow, deliberate, analytical (llms)

now you can compose both. a small model handles routing and guardrails, and the llm steps in only when needed. openinference's new decision span captures the question, choice, confidence, and probabilities, so system one calls show up right next to the reasoning they gate.

TypeSafe came out of stealth this week with Jev, the first System One Model: a frontier model that never generates prose. Unstructured state in, typed probabilistic decisions out. 70 to 500 ms, 0% type errors, calibrated probabilities on every answer, $0.042/MTok in and output free.

Here it is blocking a support draft that promised free shipping for life, at confidence 1.0.

TypeSafe came out of stealth this week with Jev, the first System One Model: a frontier model that never generates prose. Unstructured state in, typed probabilistic decisions out. 70 to 500 ms, 0% type errors, calibrated probabilities on every answer, $0.042/MTok in and output free.

When decisions get that cheap, intelligence goes into every conditional in your code. Those decisions need to be observable.

Phoenix now supports Google's Agent Development Kit for Java.

With the new OpenInference instrumentation for ADK Java, every agent invocation, LLM call, and tool execution shows up as a trace in Phoenix. Same experience the Python ADK folks have had, now for the JVM.

Phoenix now supports Google's Agent Development Kit for Java.

Phoenix now supports Meta's Muse Spark 1.3. Muse Spark 1.3 landed last week in the same 48 hours as Astra and Fable, with no blog post, and it deserves more attention than it got.

The release ships as two model IDs with identical specs. muse-spark-1.3 costs $1.25 / $4.25 per million tokens. muse-spark-1.3-contributor costs $0.10 / $0.20, roughly 12x cheaper on input and 21x on output, in exchange for letting Meta train on your prompts and completions.

Phoenix now supports Meta's Muse Spark 1.3. Muse Spark 1.3 landed last week in the same 48 hours as Astra and Fable, with no blog post, and it deserves more attention than it got.

Phoenix 20.5 is out, and the part I keep showing people is small but a huge time saver: PXI can now file the GitHub issue for the bug it just found.

• A PII Detection evaluator for conversation records, tool calls and retrieved documents included. Python and TypeScript. • Prompt versions over REST. POST a new version to an existing prompt and tag it in the same call.

Phoenix 20.5 is out, and the part I keep showing people is small but a huge time saver: PXI can now file the GitHub issue for the bug it just found.

We think MiniMax is underrated for agentic workloads, and the reason is architectural. The M-series is sparse MoE: M2 routes to ~10B of 230B parameters per forward pass, and M3 to ~23B of 428B. That is why M2 fits on 4xH100s at FP8 and why list pricing sits around $0.30/$1.20 per million tokens.

The detail we care most about from an observability standpoint: interleaved thinking. Reasoning lives in think blocks between tool calls and is meant to be kept in the message history, so every span in a trace carries the model's reasoning for that step, not just the tool call and result.

We think MiniMax is underrated for agentic workloads, and the reason is architectural. The M-series is sparse MoE: M2 routes to ~10B of 230B parameters per forward pass, and M3 to ~23B of 428B. That is why M2 fits on 4xH100s at FP8 and why list pricing sits around $0.30/$1.20 per million tokens.

Arize Phoenix now supports Z.ai and GLM-5.3. GLM-5.3 uses the same base as GLM-5.2. Every gain came from post-training. Z.ai scaled up long-horizon task environments (some representing days of real engineering work) and trained on them with their open-source slime RL framework.

The result: a 50% improvement over 5.2 on their in-house Code Bench, Terminal-Bench 3.0 from 4.6 → 28.3, DeepSWE v1.1 from 46.2 → 66.9, and open-weights SOTA on Terminal-Bench 3.0 and Agents' Last Exam. Cyber capability emerged faster than they expected.

Claude Fable 5.1 and Mythos 5.1 are the same underlying model at two safeguard levels. Fable 5.1 is generally available everywhere (API, AWS, GCP, Azure) as claude-fable-5-1, while Mythos 5.1 is restricted to vetted US organizations via cyber/life-sciences verification programs.

Long-horizon, multi-step autonomous agents is exactly the workload that's hardest to debug without observability. Trace those extended agent runs, run evals on their outputs, and track spend!

Claude Fable 5.1 and Mythos 5.1 are the same underlying model at two safeguard levels. Fable 5.1 is generally available everywhere (API, AWS, GCP, Azure) as claude-fable-5-1, while Mythos 5.1 is restricted to vetted US organizations via cyber/life-sciences verification programs.

- 55.8% on Terminal-Bench 4.0 (up from 42.0% for Fable 5) - 52.6% on Terminal-Bench-Science (up from 24.7%) - 65.0% on Humanity's Last Exam with tools.