When Model Capability Isn't the Only Selling Point, Opus 5 Changes the Game Rules

Last week, while chatting with a finance client about AI model selection, he shared an interesting detail: their team tested Anthropic's latest Opus 5 and found that scores on complex reasoning tasks improved by only 3% compared to the previous generation, yet token consumption dropped by 40% for the same tasks. He asked me if this meant Anthropic was 'milking' the product (squeezing toothpaste), or if the industry was truly undergoing a shift.