Researchers trained an OpenAI rival in half an hour for $50
Once a foundational AI model has been trained, which costs a huge amount of money, it's actually easy to clone its core functions cheaply.
Key takeaways
- The model, called s1, was refined using a small dataset of 1,000 questions and for under $50, according to TechCrunch.
- The researchers say they trained the model on just 16 Nvidia H100 GPUs.
- OpenAI’s o1 reasoning model uses a similar approach, something the buzzy AI startup DeepSeek sought to replicate with the launch of its R1 model that it claims was trained at a fraction of the cost.
Cite or link to this article
Griffin, M. (2025) 'Researchers trained an OpenAI rival in half an hour for $50', 311 Institute, 27 February. Available at: https://www.311institute.com/researchers-trained-an-openai-rival-in-half-an-hour-for-50/ (Accessed: 25 October 2026).
Fresh on the back of the release of China’s DeepSeek R1 Artificial Intelligence (AI) model which cost $5.6 Million to train, which is multiples less than other Foundation AI models, researchers in the US managed to create a low-cost AI reasoning model rivalling OpenAI’s o1 model in just 26 minutes, as outlined in a paper published last week. The model, called s1, was refined using a small dataset of 1,000 questions and for under $50, according to TechCrunch.
To do this, researchers at Stanford University and the University of Washington used a method known as distillation, which allows smaller models to draw from the answers produced by larger ones, to refine s1 using answers from Google’s AI reasoning model, Gemini 2.0 Flash Thinking Experimental. Google’s terms of service note that you can’t use Gemini’s API to “develop models that compete with” the company’s AI models.
The Future of Generative AI and AI 2040, by AI Keynote Matthew Griffin
The researchers based s1 on Qwen2.5, an open source model from Alibaba Cloud. They initially started with a pool of 59,000 questions to train the model on, but found that the larger data set didn’t offer “substantial gains” over a whittled-down set of just 1,000. The researchers say they trained the model on just 16 Nvidia H100 GPUs.
Breaking the AI Scaling Laws, Deepseek R1
The s1 model also uses a technique called test-time scaling, allowing the model to “think” for a longer amount of time before producing an answer. As noted in the paper, researchers forced the model to continue reasoning by adding “Wait” to the model’s response. “This can lead the model to doublecheck its answer, often fixing incorrect reasoning steps,” the paper says.
OpenAI’s o1 reasoning model uses a similar approach, something the buzzy AI startup DeepSeek sought to replicate with the launch of its R1 model that it claims was trained at a fraction of the cost. OpenAI has since accused DeepSeek of distilling information from its models to build a competitor, violating its terms of service. As for s1, the researchers claim that s1 “exceeds o1-preview on competition math questions by up to 27%.”
The rise of smaller and cheaper AI models threatens to upend the entire industry. They could prove that major companies like OpenAI, Microsoft, Meta, and Google don’t need to spend billions of dollars training AI, while building massive data centers filled with thousands of Nvidia GPUs.
What this means for your organisation
Copying costs far less than building.
Training a foundation model costs a huge amount, yet the core functions can be cloned cheaply. The s1 team used 1,000 questions, 16 Nvidia H100 GPUs and under $50. AI model distillation is rated Act now, with wide use expected by 2025.
Your advantage may not be the model.
The researchers found that 59,000 questions gave no substantial gains over a set of just 1,000. Smaller and cheaper systems could show that big players do not need to spend billions. Leaders should ask where value really sits, in the model or in how it is used.
Terms of service now carry real risk.
Google's terms bar using Gemini's API to develop competing models. OpenAI has accused DeepSeek of distilling its models in breach of its terms. Pair this with tech spend cost control, rated Act now, and review both what you buy and what you may legally do with it.
Three questions for your next leadership meeting
- Which of our AI costs could fall if cheaper copies match results?
- Do our vendor terms restrict how we use model outputs?
- Where does our advantage sit if the underlying model can be cloned?
The trends and technologies behind this story
From the 311 Starburst, the 311 Strategic Intelligence Unit’s map of the technologies and trends shaping the next 50 years.
Analysis by the 311 Strategic Intelligence Unit, from Matthew Griffin’s article and 311 data. First published on Fanatical Futurist.
FAQ
Why does this matter?
Once a foundational AI model has been trained, which costs a huge amount of money, it's actually easy to clone its core functions cheaply.
About the author
Matthew Griffin Founder, President and Futurist, 311 Institute
Matthew Griffin is a multi-award winning Futurist and expert in Disruption and Innovation, Geopolitics, Leadership, and Technology, who NASA have described as a "walking encyclopaedia of the future" and a "futurist Polymath."
Read full bio
Matthew Griffin is a multi-award winning Futurist and expert in Disruption and Innovation, Geopolitics, Leadership, and Technology, who NASA have described as a "walking encyclopaedia of the future" and a "futurist Polymath." 17-time best selling author of the "Codex of the Future" series, Matthew is the Founder, President and Futurist of the 311 Institute, a global Futures and Deep Futures advisory firm working with royal households, world leaders, G7, G20, and G77 governments, NGOs, and multi-national mid and mega cap firms to help them explore, shape, and lead the next 50 years of business and society.
An award-winning YouTube creator with over a million followers, with an unrivalled global reach and impact, Matthew is a highly sought-after international keynote speaker, lecturer, and mentor who collaborates with global leaders through the United Nations Alliance of Civilizations (UNAOC) and United Nations General Assembly (UNGA) to shape pivotal initiatives such as the UN’s AI for Humanity program, the United Nations Conference of the Parties (UN COP), and the World Economic Forum in Davos.
As the former Global Head of Cloud, National Security, and Enterprise Sales for companies including Atos, Dell-EMC, and IBM, Matthew has a proven track record of building multi-billion dollar business units and turning failing divisions into market leaders. His ability to identify, analyse, and communicate the implications of hundreds of emerging technologies and trends is unparalleled, and his insights are trusted by many of the world’s most respected organisations, including ABB, Accenture, Adidas, AON, ARM, BCG, Centrica, Citi, Coca-Cola, Dentons, Deloitte, Dow Jones, EY, Google, KPMG, Lego, Legal & General, LinkedIn, Microsoft, PepsiCo, Qualcomm, RWE, Samsung, Siemens AG and Siemens Energy, T-Mobile, UBS, VISA, Walmart, Workday, Worldpay and many others.
Regularly featured in the global media including the AP, BBC, Bloomberg, CNBC, Discovery, Forbes, Khaleej Times, Telegraph, TIME, ViacomCBS, WIRED, and the WSJ, Matthews mission is to help organisations create a fair and sustainable future whose benefits are shared by everyone irrespective of their ability, background, or circumstances.
What future do you need to see?
Choose one to get started on AI and intelligence and the future of your organisation.
Sources and further reading
- a paper published last week arxiv.org
- model, called s1 github.com
- according to TechCrunch techcrunch.com