Why GPT-5 used less training compute than GPT-4.5 (but GPT-6 probably won't) (epoch.ai)

OpenAI’s GPT-5 appears to have used less final “training compute” than GPT-4.5 because the team shifted effort from heavy pre‑training to much larger, more effective post‑training (reasoning/RL-style) methods. Recent research made post‑training far more efficient — historically pre‑training dominated (∼100× post‑training), but new techniques can reduce required pre‑training by roughly 10× while keeping or improving performance. Practically, that meant OpenAI could match or beat GPT‑4.5-level capabilities with a smaller final pre‑training run by doing more costly experimentation and post‑training on a smaller model; however, total development compute (experiments, data collection) likely rose, consistent with projected R&D spend growth (~$5B → ~$9B). This shift is significant because it shows a temporary change in how state‑of‑the‑art LLMs are built: marginal returns from post‑training recently outpaced further pre‑training, letting labs deliver gains faster under market/time pressure. But the authors argue this is not a new long‑term trend — post‑training scaling can’t expand indefinitely (bottlenecks in RL environments, data, and experiment compute), and tripling post‑training will soon be equivalent to tripling overall budgets. As infrastructure grows (more GPUs, new clusters), future flagships like GPT‑6 will likely return to larger pre‑training runs and higher training compute, though accounting choices (e.g., whether to include upstream synthetic‑data model costs) complicate exact comparisons.