Comet vs vLLM

Every entry in this window is educational or category content: what AI observability is, how to select a model per agentic task, which observability platforms rank in 2026, a from-scratch treatment of diffusion language models, and a build-log for a self-grading F1 radio RAG pipeline. The real Opik engineering, Agent Diagnostics for cross-trace analysis and the cost work behind it, has now fallen out of the recent window entirely. The old experiment-tracking identity is not visible at all.

Comet has completed its pivot from ML experiment tracking to LLM and agent observability, and the writing is aimed at defining the category rather than documenting releases. Cost is the recurring hook across the teaching posts, with model selection framed as a billing problem, which suggests the wedge is budget pressure rather than debugging alone. The ratio has tipped far enough that buyer education is now running well ahead of anything shipped in view.

On this evidence the feed will keep publishing category guides faster than product notes, so what Opik actually ships next is not readable here. The consistent framing of model choice as a cost decision is the one thread pointing at where the product work is going.