OpenAI May Have Really Screwed Up With GPT-4.5

OpenAI is trying to have it both ways: calling GPT-4.5 its "largest and most knowledgeable model yet" while managing expectations, cautioning that it's "not a frontier model."

Why? Probably because the company knew what the public response was going to be: muted. From AI critic Gary Marcus' calling the LLM is a "nothingburger" to an anonymous expert who told Ars Technica that the model is a "lemon," it appears that the amply-hyped new model is seriously lacking in the type of juice that made the original ChatGPT, or its followup GPT-4, become enormous cultural and financial touchstones.

Priced at a whopping $75 per million input tokens and $150 per million output tokens, GPT-4.5 is 30 times more expensive than GPT-4o, OpenAI's "reasoning" model — and beyond the sticker shock, it's apparently quite slow and has a focus on "emotional intelligence" instead of more traditionally analytic tasks.

While haters will inevitably hate, and the hype around OpenAI's new models has been decreasing steadily, the sense of underwhelmingness this time is pervasive.

Marcus pointed out that GPT-4.5 still hallucinates and makes errors — and even OpenAI admitted, per the MIT Technology Review, that it still made stuff up 37 percent of the time in a key test. The anonymous expert who spoke to Ars, meanwhile, blasted the model's massively increased price and said its performance doesn't bear out such a markup.

Many of OpenAI's self-serving accolades for GPT-4.5 hinge on its ability to chat with sensitivity and intuitiveness — qualities that the company thinks people want, apparently.

"It has the ability to engage in warm, intuitive, natural, flowing conversations," OpenAI research scientist Mia Glease told the MIT Technology Review. "And we think that it has a stronger understanding of what users mean, especially when their expectations are more implicit, leading to nuanced and thoughtful responses."

Basically, the pitch seems to be that the model has been given AI sensitivity training to help it chat with humans more naturally — leading observers online to clown on GPT-4.5 for being emotionally intelligent but not that good at its job.

"So GPT-4.5 is the model you vent to?" one user quipped. "I want a smart model, not a model that understands my feelings."

Indeed, OpenAI chief research officer Mark Chen admitted in a recent interview that the benchmarks used to track how much the model had advanced are "vibes-based" — a wild acknowledgment, seemingly, of the company's mentality surrounding its hot new model.

More on OpenAI: Sam Altman Says OpenAI Has Run Out of GPUs