Meet s1 – The AI Model Trained for Less Than $50, Challenging Tech Giants

Artificial intelligence research has long been dominated by tech giants with vast financial resources. However, a new study by researchers at Stanford University and the University of Washington has shaken up this narrative. The team successfully trained an AI reasoning model, called s1, for less than $50 in cloud computing credits—a fraction of the cost typically associated with training advanced AI models.

This groundbreaking achievement raises important questions about the commoditization of AI and the barriers to entry in the field. If a small team of researchers can replicate the reasoning abilities of multi-million-dollar AI models with minimal resources, what does this mean for the future of AI development?

Meet s1 – The AI Model Trained for Less Than $50, Challenging Tech Giants

The s1 model has demonstrated impressive capabilities, performing on par with leading AI reasoning models like OpenAI’s o1 and DeepSeek’s R1 in math and coding benchmarks. The researchers have made s1 publicly available on GitHub, along with the dataset and training code, allowing others to replicate or build upon their work.

They created the model using an off-the-shelf AI model from Qwen, an AI lab owned by Alibaba. The researchers fine-tuned this base model through a process called distillation, a technique that extracts reasoning capabilities from a more advanced model by training on its answers. In this case, the researchers distilled s1 from Google’s Gemini 2.0 Flash Thinking Experimental model.

Distillation has become an increasingly popular method for training AI models at a lower cost. Some researchers at Berkeley also used a similar approach and created an AI reasoning model for approximately $450. However, the fact that Stanford and Washington researchers managed to achieve similar results for just $50 is particularly significant.

Challenges for Big AI Labs

The success of s1 highlights an uncomfortable reality for major AI companies: the cost of replicating state-of-the-art AI models is dropping rapidly. If a small team with limited resources can closely match the performance of multi-million-dollar AI models, where does that leave the companies investing billions in AI development?

Unsurprisingly, big AI labs are not happy. OpenAI, for instance, has accused DeepSeek of improperly harvesting data from its API to train its own reasoning model, R1. Similarly, Google forbids reverse-engineering its models to develop competing AI products. While Google provides free access to its Gemini 2.0 Flash Thinking Experimental model through Google AI Studio, it enforces daily usage limits to prevent large-scale data extraction.

We have reached out to Google for comment regarding s1’s development but have yet to receive a response.

How Was s1 Trained for Just $50?

The researchers designed a small but highly effective training dataset of 1,000 carefully curated questions, each paired with detailed answers and step-by-step reasoning provided by Google’s Gemini model.

Using 16 Nvidia H100 GPUs, they trained s1 in under 30 minutes. One of the researchers, Niklas Muennighoff, estimated that renting the required cloud computer today would cost just $20. “This demonstrates that researchers can efficiently distil AI models with minimal computational power.”

To further enhance s1’s reasoning, the team used a simple but effective trick: they instructed the model to “wait” before answering a question. This small tweak gave the AI more time to process information, leading to slightly more accurate responses.

What This Means for the Future of AI

As AI technology advances, major companies like Meta, Google, and Microsoft are planning to invest hundreds of billions of dollars in AI infrastructure by 2025. This investment will support the development of next-generation AI models, but s1’s success proves that smaller research teams can still make significant contributions without vast financial backing.

While distillation is a cost-effective way to replicate existing AI capabilities, it does not necessarily push the boundaries of AI innovation. Truly groundbreaking advancements may still require large-scale investments and powerful infrastructure. However, as AI models become easier and cheaper to replicate, the AI industry could shift towards a more open and competitive landscape.

The success of s1 is a wake-up call to the AI world—cutting-edge AI is no longer reserved for tech giants.

See also: Generative AI vs. Agentic AI: Decoding the Future of Artificial Intelligence