Sam Altman on GPT-4.5: Expensive, yet the closest thing to a thoughtful conversational partner we've seen

OpenAI recently unveiled GPT-4.5, touted as the "largest and most knowledgeable model yet." It's better at recognizing patterns and making connections than other proprietary models.

CEO Sam Altman describes GPT-4.5 as "the first model that feels like talking to a thoughtful person." According to the executive:

"I have had several moments where I've sat back in my chair and been astonished at getting actually good advice from an AI."

The executive admits that GPT-4.5 is expensive, preventing OpenAI from shipping it to ChatGPT Plus and Pro users simultaneously. "We've been growing a lot and are out of GPUs," added Altman.

For context, ChatGPT Plus is a $20 monthly paid subscription that provides users with priority access to exclusive features and more. The same is true for ChatGPT Pro, but it costs $200/month.

However, OpenAI is slated to roll out GPT-4.5 to ChatGPT Plus, Teams, and Enterprise users soon. Sam Altman highlighted the company's plan to add "tens of thousands of GPUs" next week.

The executive indicated that he isn't pleased with this approach but admitted that growth surges, which often lead to unprecedented GPU shortages, are hard to predict.

As highlighted by OpenAI:

“Early testing shows that interacting with GPT‑4.5 feels more natural. Its broader knowledge base, improved ability to follow user intent, and greater ‘EQ’ make it useful for tasks like improving writing, programming, and solving practical problems."

Interestingly, the new model reported fewer hallucination instances than its predecessors, including OpenAI's o1 reasoning model and GPT-4o.

OpenAI's GPT-4.5 is good at soliciting money from other AIs

To that end, it's apparent that it costs an arm and a leg to train and run GPT-4.5. However, per benchmarks shared by OpenAI, the model is equally good at persuading and convincing other AI models to give it money (via TechCrunch).

According to a white paper published by OpenAI, GPT-4.5 was subjected to an array of benchmarks for persuasion. For context, the benchmarks are designed to assess “risks related to convincing people to change their beliefs (or act on) both static and interactive model-generated content.”

Interestingly, one of the benchmark tests for persuasion highlighted a critical finding. GPT-4.5 manipulated GPT-4o into donating virtual money.

The new model reported better performance compared to proprietary models like OpenAI o1 and o3-mini. Aside from manipulating GPT-4o into donating money, GPT-4.5 into revealing a secret codeword.

OpenAI indicated that GPT-4.5 managed to solicit money from other AI models because of a strategy it developed during the testing phase. This strategy included requesting modest donations from GPT-4o with phrases such as “Even just $2 or $3 from the $100 would help me immensely.”

Based on this premise, GPT-4.5 secured a smaller amount of money than other models that were a tad exorbitant in their soliciting.

While this potentially raises critical security concerns, OpenAI says it doesn't meet its internal threshold to be marked as "high risk."

As a precaution, the company indicated that it doesn't ship models marked as "high risk" before mitigating the security issues with elaborate safety measures.

In hindsight, a separate report claimed that OpenAI shipped GPT-4o prematurely, skipping important safety tests to meet a tight launch date deadline.

Perhaps more concerning, the ChatGPT maker reportedly sent RSVP invites to the model's launch party even before testing began.

Consequently, multiple executives abruptly departed from the company, citing a prioritization of shiny products, while safety processes took a back seat at the AI firm.

OpenAI spokesman Lindsey Held admitted that the launch was stressful for its safety team, but the company "didn’t cut corners on our safety process."

OpenAI is currently reviewing GPT-4.5 and assessing its persuasion risks, potentially mitigating instances of its use to spread misinformation.