OpenAI prepares to launch GPT-5, but big leaps are unlikely

Internal testing shows GPT-5 edges ahead of GPT-4—better code, cleaner math, sharper step-by-step thinking. But no breakthrough. No leap. OpenAI even scrapped “Orion,” the original GPT-5 push, and settled on GPT-4.5 instead. Translation: scaling Transformers is hitting a wall.

System pivot: OpenAI’s now betting on hybrid models and reinforcement-trained reasoning to chase agent-grade AI. The age of the solo LLM may be winding down.