GPT-4.5 = Big Model Energy

After GPT-4 arrived in 2023, expectations grew for GPT-5, while rumors about projects called Strawberry and Orion circulated. OpenAI released o1 in December, introducing a reasoning-focused model that outperformed GPT-4 on math, coding, and logic. Orion became GPT-4.5, a much larger model built by scaling pretraining and post-training.

OpenAI researchers describe GPT-4.5 as better at understanding what people mean and having nuanced conversations. It scored 61.9% on SimpleQA, compared with GPT-4o’s 38.4%, and its reported hallucination rate was 37%, versus 61.2% for GPT-4o. The presenter says it also performs well at writing, storytelling, humor, and brainstorming.

The presenter says GPT-4.5 performed better on softer tasks in internal testing, with some evaluation relying on human feedback and subjective judgment. Researchers asked testers where it improved or regressed and used their feedback to guide alignment. Because writing quality and emotional intelligence are difficult to measure consistently, OpenAI wants public users to try the model and share their impressions.

GPT-4.5 costs 30 times more per input token and 15 times more per output token than GPT-4o, making it difficult to justify for large-scale deployment. It also trails reasoning-focused models such as o1 on complex STEM, advanced math, and difficult coding tasks. The presenter argues that broad pretraining and inference-time reasoning can complement each other, and describes GPT-4.5 as a bridge toward models that combine both.