OpenAI Unveils GPT-4.5, But Faces GPU Shortages
OpenAI introduced its latest AI model, GPT-4.5, today, but the rollout encountered unexpected hurdles.
CEO Sam Altman revealed that the company ran out of GPUs—critical computing resources—just before the launch. As a result, access to GPT-4.5 is currently restricted to ChatGPT Pro users ($200 per month) and developers using the paid API, who will now pay $75 per million tokens, a sharp increase from the previous $15 rate for GPT-4.
Initially, OpenAI had planned to make the model available on the more affordable ChatGPT Plus plan ($20 per month), where it typically debuts new releases.
“It’s a massive and costly model,” Altman explained. “Our intention was to launch it for both Plus and Pro users simultaneously, but due to rapid growth, we’ve hit a GPU bottleneck. This isn’t ideal, but predicting demand surges precisely is challenging.”
To address the shortage, OpenAI expects to acquire “tens of thousands” of additional GPUs in the coming week, enabling wider access for Plus and Team subscribers, followed by Enterprise and Edu users shortly after. Altman assured that “hundreds of thousands” more GPUs are on the way, anticipating heavy usage.
These resources will be essential not only for GPT-4.5 but also for the upcoming GPT-5 release, expected in the next few months. Additionally, OpenAI remains committed to investing $500 billion in AI infrastructure over the next four years as part of Project Stargate.
While GPT-4.5 represents an advancement, Altman cautioned against expecting dramatic improvements in benchmark performance, as there is no universal metric for measuring its enhancements. “It offers a unique kind of intelligence, something almost magical. I’m excited for people to experience it firsthand.”
During a livestream, OpenAI engineers discussed how the company assesses the model’s human-like qualities, emphasizing its “vibes” and “creative intelligence.” This suggests the AI could be optimized for roles where emotional intelligence and user interaction matter more than raw computational ability.
“When we refer to ‘vibes,’ we’re talking about the model’s emotional intelligence—how collaborative, warm, and engaging it feels,” an engineer explained. “We evaluate this using curated prompts and select trainers who best align with our standards.”
Newer Articles
- Kia’s EV4 Electric Sedan Set to Arrive in the US by 2026
- What’s Streaming on Paramount+ With Showtime in March 2025
- Will the iPhone 17 Be Apple’s Boldest Upgrade Yet?
