Google’s Gemini 4 Argon turns inputs such as photographs into finished artifacts through tools and subagents, according to Google’s curated demonstrations, while Gemini 4 Carbon targets longer engineering and cybersecurity work, where small gains in per-step reliability compound across a whole task. Mistral’s Timestral brings pretrained, zero-shot forecasting to numerical series, with quantile outputs that express uncertainty. The article covers how to evaluate an Argon-based system on completed projects, why duration and an external signal matter for Carbon-style agents, and how Timestral handles multiple series, future-known information, and synthetic pretraining. It closes with a proposed design that pairs a numerical forecaster with a generalist agent, using clear, checkable interfaces between them. The article frames the contrast through two jobs: turning photographs of a desk into a playable racing game, which requires visual understanding, code, tools, and repeated correction, and predicting tomorrow’s electricity demand, which requires learning how numerical signals evolve, including how uncertain their futures are. Both involve prediction but demand very different machinery. The article’s reading of the week is that useful AI increasingly depends on matching the model, its inputs, and its feedback to the job.