Researchers Build AI Model For $50 That Rivals OpenAI And DeepSeek

A group of AI researchers from Stanford and the University of Washington have trained a high-performing AI reasoning model for less than $50 in cloud computing costs, according to a research paper released last Friday, as first reported by TechCrunch.

The model, called s1, performs on par with advanced reasoning AIs like OpenAI’s o1 and DeepSeek’s R1 in math and coding tests. It is freely available on GitHub, along with the training data and code.

To develop s1, the researchers started with a pre-existing AI model and fine-tuned it using a process called distillation. This method extracts reasoning skills from a more advanced AI by training on its answers.

The team revealed that s1 was distilled from Google’s Gemini 2.0 Flash Thinking Experimental model.

The group of researchers has built the open-source AI model that rivals OpenAI’s o1-preview in solving tricky math and logic problems. Their secret? A simple trick that gives AI more time to think before answering.

The method, called “budget forcing,” works by making the AI take extra steps when solving problems instead of rushing to an answer. By giving itself more time, the AI can double-check its work and improve accuracy.

The team trained their model, called s1-32B, using just 1,000 carefully chosen example questions with step-by-step explanations.

Despite this small dataset, the model outperformed OpenAI’s o1-preview in complex math tests, such as MATH and AIME24, improving results by up to 27%. It even managed to boost its own score on a test from 50% to 57% just by thinking longer before finalizing answers—without any extra training.

This is a big deal because most AI improvements rely on huge amounts of new training data. Instead, this research suggests that many AI models already have strong reasoning skills hidden inside them—they just need the right techniques to unlock them.

The study also tested different ways to extend an AI’s thinking time. One method involved letting the AI refine its answers step by step, while another had it generate multiple possible answers at once and pick the best one. A combination of both, using a tree-like decision process, worked best.

While this technique has limits—such as the AI’s memory space for calculations—the researchers believe further tweaks could push its abilities even further. They also suggest that reinforcement learning, a method where AI learns from trial and error, could make test-time thinking even smarter.

By making their model and research freely available, the team hopes to encourage more open and transparent AI development, helping others build smarter and more reliable AI systems.