OpenAI has launched its largest artificial intelligence model with fewer “hallucinations”, the latest in a flurry of AI releases by US tech groups.
The San Francisco-based company on Thursday unveiled GPT-4.5, its long-awaited update to the technology that underpins its popular product ChatGPT. In early tests, its hallucination rate, where AI systems generate inaccurate information, was 37 per cent compared with nearly 60 per cent on its predecessor GPT-4o.
With GPT-4.5, OpenAI is continuing to bet on big, expensive large language models despite the advent of highly capable smaller products, such as Chinese start-up DeepSeek’s R1, which are open, cheaper and more accessible for developers.
It comes as competition is growing within the fast-developing AI industry as tech groups have rushed to launch their latest models in recent weeks. Anthropic revealed its Claude 3.7 Sonnet on Monday, which followed last week’s launch of Grok 3, the latest model from Elon Musk’s xAI.
OpenAI in a blog post on Thursday said GPT-4.5 had “broader knowledge and a deeper understanding of the world, leading to reduced hallucinations and more reliability across a wide range of topics”.
“With every new order of magnitude of compute comes novel capabilities,” the company said, adding GPT-4.5 was “at the frontier of what is possible in unsupervised learning”.
OpenAI has been at the forefront of a global race to lead the AI industry, raising tens of billions from investors to fund bigger models with increased capabilities that require vast amounts of computing power.
It is in talks with SoftBank and other investors to raise up to $40bn at a valuation of $300bn, including the new money. Anthropic is also fundraising about $3.5bn at a $60bn-plus valuation, said two people with knowledge of that process.
However, the huge costs of running larger models have led OpenAI to consider withdrawing developer access to GPT-4.5 amid fierce competition on pricing from rivals.
OpenAI said while a preview of GPT-4.5 will be made available to developers who pay to use OpenAI’s models through its application programming interface (API), this access could be revoked in the future.
AI groups largely generate revenue through paid API access and individual subscriptions. OpenAI said it will see how developers use the powerful model and whether it is worth offering to them considering the high cost of running it.
The company said: “GPT‐4.5 is a very large and compute-intensive model, making it more expensive than and not a replacement for GPT‐4o [its predecessor]. Because of this, we’re evaluating whether to continue serving it in the API long-term as we balance supporting current capabilities with building future models.”
OpenAI chief executive Sam Altman has previously said GPT-4 cost more than $100mn to train, and such costs are widely expected to increase as the size and capabilities of models scale and require more computing power to train and run.
In a post on X after the announcement, Altman said the company was “out of GPUs”, the chips required to run and train AI systems.
“This isn’t how we want to operate, but it’s hard to perfectly predict growth surges that lead to GPU shortages,” he said, adding they expected to get more in the coming weeks.
He also highlighted GPT-4.5 was not focused on reasoning and would not beat industry benchmarks but it had “a different kind of intelligence and there’s a magic to it I haven’t felt before”.
Additional reporting by George Hammond in San Francisco and Melissa Heikkilä in London
