GPT-4.5 = Big Model Energy | YC Decoded

YC’s Tom Blomfield breaks down GPT-4.5: OpenAI’s largest model yet excels at emotional intelligence but costs 30x more than GPT-4o.

• GPT-4.5 is potentially 10x larger than GPT-4 and represents a scaling step in pre-training, not a reasoning model.
• Hallucination rate drops from 61.2% (GPT-4o) to ~37%; SimpleQA accuracy jumps from 38.4% to 61.9%.
• Cost is prohibitive: 30x more expensive per input token and 15x per output token than GPT-4o.
• GPT-4.5 underperforms specialized reasoning models (o1, o3) on math, coding, and structured STEM tasks.
• OpenAI used subjective ‘vibes testing’ with human trainers to evaluate writing quality and emotional intelligence.
• Sam Altman has signaled GPT-5 will converge pre-training and reasoning into a unified architecture.
• The model is positioned as a bridge — broad world knowledge plus emotional nuance, awaiting reasoning integration.

2025-03-11 · Watch on YouTube