Ep. 55 — The harness, not the model: a wrapper takes Opus 5 from 30% to a perfect score, a 27-billion-parameter agent beats the frontier at replicating research, and nobody knows who made the week's best free model

The harness, not the model: a wrapper takes Opus 5 from 30% to a perfect score, a 27-billion-parameter agent beats the frontier at replicating research, and nobody knows who made the week's best free model 00:48 — Nvidia's research puts Claude Opus 5 at 100% on an interactive reasoning benchmark - by changing the harness, not the model 02:17 — Inherent's Faraday - a 27-billion-parameter model wrapped in the right scaffolding - says it beat Claude Opus 4.8 and GPT-5.5 at replicating published research 03:47 — The builder question: an anonymous free model tops developers' lists while a hobbyist's un-guardrailed agent gets him banned from Resy Daily Prompt with Archer & Iris.