Are rising AI training costs producing only narrow capability gains?
Core argument: Composite benchmarks drawn from broader, cleaner task sets place Gemini roughly in line with GPT-5-class models, with any overall gap small in either direction.
Composite benchmarks built from broader, cleaner sets show Gemini roughly in line with GPT-5-class models. On the strictest reasoning tasks—low-leakage exams—Gemini almost certainly does not lead. In many cases it likely trails. Either way. whatever gap exists overall, if any, is small. The Gemini release fits the same trend seen across the industry: growing training costs and shrinking capability gains. Progress has slowed on the tasks that matter most for general reasoning. The gains that do appear come from domains still responsive to brute-force scaling—vision, video, and certain coding workflows—not from breakthroughs in abstraction, logic, or planning. If Gemini had made a clear leap in those harder areas, the missing benchmarks would have been on the above list. They were not, almost certainly because the gains are not there. The core pattern remains: large increases in compute and engineering produce modest, targeted improvements rather than broad advances in general capability.