The GPT-4 wall was the phenomenon that large language models available publicly between early 2023 (GPT-4's public release) and 2025 were in most ways not significantly more intelligent than GPT-4-0314 itself. Despite massive advancements in cost optimization, much wider availability of GPT-4-level models, inference-time scaling, and better median user preference tuning, frontier models during the wall period were not significantly better at certain hard tasks, and could sometimes be seen as downgrades relative to GPT-4.