Claude Opus 5.5 and GPT-6 Sol/Luna launched the same day, both cutting prices 40-50% while landing within a point of each other on independent coding evals — a sign that frontier-model differentiation is shifting from raw capability to cost and speed. For DotTif Labs, where cost per request already shapes model choice, this is the point where routing between providers starts to pay for itself rather than being a hedge.
A Pentagon review found human operators over-trusted an AI targeting recommendation, contributing to a strike that hit a school — the most consequential documented case yet of automation bias in a high-stakes decision-support system. It's the sharpest evidence so far that the failure mode with copilot-style AI tools often isn't the model being wrong, it's the human's skepticism running out too early.
via bloomberg via hn
Within days of launch, Jev's 'decision model' idea already has six independent clones, a reproducible benchmark, and a 25-line reimplementation proving the core trick is trivial to copy — a live demonstration of how fast an idea, not just a model, can commoditize. If you're adding a routing or decision layer to an agent pipeline, this is the pattern worth watching before you build one from scratch.
Why read anyway — The original walks through the actual 25 lines of code, which makes concrete exactly why this pattern is so easy to reproduce — something a summary bullet can't fully convey. original ↗