OpenAI has outlined its plans for GPT-4.5, but is struggling with a shortage of GPUs.
The launch of GPT-4.5 had been coming for a while, but OpenAI’s CEO Sam Altman says the rollout will be in phases as the company “runs out of GPUs.” ChatGPT Pro subscribers will get access first, followed by Plus users next week.
Most expensive AI model to date
GPT-4.5, under the name Orion, is OpenAI’s largest and most expensive model to date. It requires enormous computing power: Altman speaks at X of “tens of thousands of additional GPUs” needed to enable wider access. The cost is not minus: $75 per million input tokens (about 750,000 words) and $150 per million output tokens. The prices are much more expensive than GPT-4o, as much as 30 times and 15 times higher. GPT-4o costs $2.50 and $10 for the same number of tokens, respectively. OpenAI is not even sure if GPT-4.5 is sustainable in the long term within its API offerings.
Better, but not revolutionary
GPT-4.5 outperforms GPT-4o on skills such as factual questions(SimpleQA) and creative tasks, but lags on benchmarks for mathematical and reasoning questions. Models such as DeepSeek R1 and Claude 3.7 Sonnet score better here. Still, OpenAI claims in a white paper that GPT-4.5 is “warmer, more natural and better at intention understanding” than previous models.
OpenAI’s biggest challenge is that the benefits of traditional AI training are diminishing. The industry is increasingly focusing on reasoning models, which solve problems step by step rather than just predict.