“It’s a lemon”—OpenAI’s largest AI model ever arrives to mixed reviews
1 min read
Summary
OpenAI’s new traditional AI model, GPT-4.5, has been criticised for its poor performance despite its enormous size and high cost.
Compared to GPT-4o, the new model provides only marginally better performance at a cost that is 30 times higher for the input and 15 times higher for the output.
This seemingly confirms rumours of diminishing returns in relation to training unsupervised learning LLMs.
When compared to OpenAI’s simulated reasoning models, including o1 and o3, the model also underperforms dramatically.
OpenAI has acknowledged the model’s limitations and is currently evaluating whether to keep it in API mode long-term.
The research launch has mostly been characterised as a low-key event for ChatGPT Pro users.
