• OpenAI released GPT-4.5 on Thursday.
• The model is designed to be more general-purpose than OpenAI's STEM-focused reasoning models.
• OpenAI says it's best for "tasks like writing, programming, and solving practical problems."

On Thursday, OpenAI released what it claims to be its largest and most powerful model to date: GPT-4.5.

OpenAI CEO Sam Altman described it in a post on X on Thursday as "the first model that feels like talking to a thoughtful person."

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

Altman added in his post that the model will be "giant" and "expensive." And while OpenAI "really wanted to launch it to plus and pro at the same time — " the company's paid subscription tiers — it simply ran out of GPUs, he explained.

"We will add tens of thousands of GPUs next week and roll it out to the plus tier then," he said.

Silicon Valley has been at odds about the best way to make AI models smarter and more powerful. But GPT-4.5 makes a case for the conventional wisdom that the more data and computational resources that go into a model, the better it becomes.

OpenAI's chief research officer, Mark Chen, told the newsletter Big Technology that the company has not yet seen diminishing returns from scaling.

"We're very rigorous about how we do this," Chen said. "We make projections based on all the models we've trained before on what performance to expect, and in this case, we put together the scaling machinery, and this is the point that lies at that next order of magnitude."

And while training costs remain high, OpenAI has found less expensive ways to run increasingly big models. Inference costs "have dropped many orders of magnitude since we first launched GPT-4," Chen told Big Technology.

The basics

On Thursday, the company released GPT-4.5 in a research preview to users who pay $200 a month for ChatGPT Pro and developers in the API. Next week, OpenAI aims to bring it to ChatGPT Plus, Team, and Edu users.

In a livestream demonstration of GPT-4.5's abilities on Thursday, Amelia Glaese, a member of OpenAI's technical staff, said GPT-4.5 is the latest advancement of OpenAI's "unsupervised learning" paradigm which focuses on scaling up models on "word knowledge, intuition, and reducing hallucinations."

Meanwhile, its o1 series of reasoning models, which it released last year, are designed to think before responding and are better suited for quantitative tasks.

It picks up better on social cues

In practice, GPT-4.5 is the most natural conversationalist and emotionally intelligent of OpenAI's models. It responds more adeptly to social cues than OpenAI's STEM-focused reasoning model, o1, as a function of its greater knowledge base and stronger contextual understanding.

Raphael Lopes, a member of OpenAI's technical staff, demonstrated how GPT-4.5 would reframe an angry text to a friend with more tact than o1.

It's trained on "vibes"

GPT-4.5 is aligned to be a "better collaborator" so conversations with it feel "warmer, more intuitive and emotionally nuanced," Lopes said. OpenAI tested GPT-4.5 against 4o, a multimodal model it released in May, on a "vibes" test set that measures creative intelligence and emotional intelligence.

It's less prone to hallucinate

GPT-4.5 outperforms other models in accuracy and produces significantly fewer hallucinations, the company said.

The model's "knowledge base, stronger alignment with user intent, and improved emotional intelligence make it well-suited for tasks like writing, programming, and solving practical problems," OpenAI said in the GPT-4.5 System Card published on Thursday.

OpenAI did not immediately respond to Business Insider's request for a comment.

• Chinese startup DeepSeek shocked markets this week after releasing a cheaper rival to OpenAI's o1.
• Silicon Valley has reacted to DeepSeek's release with a mix of panic and awe.
• Some AI startups see an opportunity in DeepSeek's open-source success.

In the tech industry, the tides can turn quickly, especially when it comes to AI.

Last week, OpenAI was the industry leader, developing what many saw as the most advanced AI models on the market, which led to a skyrocketing valuation.

This week, its standing was in question as Silicon Valley eyed a more cost-effective competitor: DeepSeek.

The Chinese company recently released a challenger to OpenAI's o1 reasoning model called R1. Users who've tested both said R1 rivals the capabilities of o1 and comes at a substantially cheaper cost.

The news shocked markets on Monday, leading to a stock sell-off that wiped almost $1 trillion in market cap. AI insiders said the frenzy is warranted: DeepSeek's methods are a game changer for the industry.

CEOs of startup companies facilitating the AI boom by supplying hardware, security services, and building agents told Business Insider that DeepSeek's success creates more opportunities for smaller companies to flourish.

Roi Ginat, the cofounder and CEO of EndlessAI, which develops the video AI assistant Lloyd, said DeepSeek's success could widen the pool of who can develop AI technology — and who can access it.

"DeepSeek's success represents a democratization of AI development, where smaller teams with limited resources can meaningfully compete with well-funded tech giants," Ginat wrote by email. "This has catalyzed a wave of innovation from startups and research labs previously considered peripheral to the field."

While OpenAI might not lose its standing in the industry, Ginat said its role could change. "The industry is witnessing a fascinating tension between two competing visions. One focuses on pursuing artificial general intelligence (AGI) through increasingly powerful and comprehensive models. The other emphasizes practical applications through efficient models and methods targeted at specific use cases and benchmarks," he said, comparing OpenAI and DeepSeek. "This tension drives innovation in both directions, and also exists within the big companies."

Pukar Hamal, the CEO of SecurityPal, which helps companies like OpenAI complete security questionnaires, said the industry should temper expectations of immediate change.

"If the DeepSeek team truly can cut training and inference costs by an order of magnitude, it could spark far broader deployment of AI than analysts anticipate," Hamal, told Business Insider. "On the flip side, it'll take more than a few tough earnings calls to make the biggest AI players reconsider the staggering GPU investments we're seeing for 2025."

Meta recently committed $60 billion to AI infrastructure investments. President Donald Trump also announced Stargate last month, a joint venture between OpenAI, Oracle, and SoftBank that will invest $500 billion into AI infrastructure across the country.

One of the biggest debates among AI innovators is whether open-source models, which the public can access and modify, are more likely to drive breakthroughs than closed-source models. OpenAI says it keeps its models closed for safety, while DeepSeek's models are open-source.

Satya Nitta, the cofounder and CEO of Emergence AI, a company developing AI agents, said that "DeepSeek R1 is a meaningful advance in broadening access to AI reasoning, spotlighting the power of open source and setting a new benchmark for reasoning."

Hamal said we should still approach open-source development cautiously — even if it'll eventually dominate the industry.

"An 'open source' model of unknown alignment invites serious public safety and regulatory questions. If DeepSeek's mobile app keeps climbing the charts, we could end up with a discussion similar to the recent calls to block TikTok in the US," he said. White House advisor David Sacks also raised concerns about DeepSeek's training methods when he told Fox News that it is 'possible' DeepSeek used OpenAI's models to train its own AI model.

Still, "openness typically wins in the long run," Hamal said. "If DeepSeek helps reset an increasingly closed foundational model market, that can be a net positive — so long as we maintain the guardrails that protect customers and the public at large."

If there's one lesson AI executives are taking away from this week, though, it's that it's possible to do more with fewer resources.

Matthew Putman, CEO of Nanotronics, which designs AI-controlled factories, said, "To me, the competition itself is less significant than the validation of a broader principle: AI models can be built more affordably and applied far beyond large language models."