Researchers from Stanford and the University of Washington have trained an AI “reasoning” model for less than $50 in cloud compute credits. Their model, s1, rivals OpenAI’s o1 and DeepSeek’s R1 in math and coding benchmarks, raising serious questions about AI commoditization.
Now available on GitHub, s1 was distilled from Google’s Gemini 2.0 Flash Thinking Experimental model. Distillation—a technique that extracts reasoning capabilities from a more powerful AI by training on its outputs—is proving to be a cost-effective way to replicate cutting-edge AI. Last month, Berkeley researchers used a similar process to build a competing model for just $450.
Unlike large-scale reinforcement learning, s1’s creators used supervised fine-tuning (SFT) on a dataset of just 1,000 carefully curated questions. The training—running on 16 Nvidia H100 GPUs—took less than 30 minutes. Stanford researcher Niklas Muennighoff estimates that renting that compute today would cost about $20.
This raises an unavoidable question: If advanced AI models can be cloned for pocket change, is there a moat for AI giants investing billions? OpenAI has already accused DeepSeek of harvesting its API data for model distillation—which is ironic, considering OpenAI scraped the entire public web without permission. Meanwhile, Google provides free access to Gemini 2.0 Flash Thinking Experimental via AI Studio, but prohibits reverse engineering for competitive use.
Framing AI’s future as a choice between algorithmic efficiency and brute force compute is a false dichotomy—AI’s future will require both. One thing is certain: AI innovation is no longer exclusive to those with deep pockets.
Author’s note: This is not a sponsored post. I am the author of this article and it expresses my own opinions. I am not, nor is my company, receiving compensation for it. This work was created with the assistance of various generative AI models.