OpenAI’s $19.6 Billion Pivot As GPT-4 Reaches Its Limits

Two years after GPT-4’s launch, the world is still waiting for its successor—and wondering what’s causing the delay.

OpenAI reportedly struggled to achieve significant performance gains over GPT-4, signaling that its pre-training methods were approaching diminishing returns.

Earlier this month, OpenAI CEO Sam Altman hinted in a tweet that GPT-4.5 and GPT-5 would be released within “weeks/months,” but he remained tight-lipped about the reasons behind GPT-5’s extended development timeline, Fortune’s Data Sheet reported.

However, the bigger revelation is OpenAI’s quiet shift in strategy.

Since 2018, the company has relied on a scaling approach—building ever-larger models and feeding them increasing amounts of data during pre-training (the "P" in GPT stands for “pre-trained”).

But that approach has now hit its limits.

Two former OpenAI employees told Fortune that “Orion,” the internal name for GPT-4.5 and the last of OpenAI’s non-chain-of-thought models, was originally intended to be GPT-5.

However, OpenAI reportedly struggled to achieve significant performance gains over GPT-4, signaling that its pre-training methods were approaching diminishing returns.

The upcoming GPT-5 will need to rival Anthropic’s Claude 3.7 Sonnet hybrid AI model, which blends fast, intuitive reasoning with a slower, more deliberate approach.

Some believe GPT-5 could be a major step toward artificial general intelligence (AGI), though several experts argue that GPT-style models will never fully match human cognitive abilities.