(无标题)
Fifth, for the sake of specificity re: the test-time compute point, I predict that GPT-5 with high test-time compute settings (e.g. scaffolding/COT lengths etc. equivalent to the engineer market rates mentioned here) will be above the trend-line.