Stanford and UW Researchers Create AI Reasoning Model Rivaling OpenAI's o1 for Under $50

US researchers built a DeepSeek competitor for less than a tank of gas - and it's actually good

TL;DR: AI researchers from Standard and the University of Washington developed a competitive low-cost AI model, s1, using a small dataset and a budget under $50. AI researchers from Standard and the University of Washington claim to have made significant progress in the development of low-cost AI models. Published in a recent research paper, the model entitled 's1' was reportedly built using a small dataset of 1,000 questions and a budget of less than $50. Stanford's AI Research Lab (Credit: Flickr) The development was achieved through a process called distillation. Distillation allows smaller models to leverage the capabilities of larger models throughout the training process. In this instance, the s1 model was distilled from Google's Gemini 2.0 - utilizing the 'thinking' process behind each answer from Gemini Flash 2.0 experimental. Google's terms of service prohibit using Gemini's API to develop models that compete with their AI models, leaving s1 in somewhat of a legal gray area. No official comments have been made in response to the development. The s1 model reportedly rivals the coding and mathematics performance of OpenAI's o1 and DeepSeek's r1, achieving strong performance in benchmark results. While it does not surpass the industry-leading models, it comes surprisingly close considering its budget. To put things in perspective, s1 won't be shattering markets in the way DeepSeek's r1 did. However, it does have strong implications for AI firms' business models. Ultra-low-cost training proves that models can be developed without billions of dollars of compute power, effectively showing that the 'moat' between smaller players and the giants may be beginning to close.

Researchers create reasoning model for under $50, performs similar to OpenAI's o1

Why it matters: Everyone's coming up with new and innovative ways to work around the massive costs involved with training and creating new AI models. After DeepSeek's impressive debut, which shook Silicon Valley, a group of researchers has developed an open rival that reportedly matches the reasoning abilities of OpenAI's o1. Stanford and University of Washington researchers devised a technique to create a new AI model dubbed "s1." They have already open-sourced it on GitHub, along with the code and data used to build it. A paper published last Friday explained how the team achieved these results through clever technical tricks. Rather than training a reasoning model from scratch, an expensive endeavor costing millions, they took an existing off-the-shelf language model and "fine-tuned" it using distillation. They extracted the reasoning capabilities from one of Google's AI models - specifically, Gemini 2.0 Flash Thinking Experimental. They then trained the base model to mimic its step-by-step problem-solving process on a small dataset. Others have used this approach before. In fact, distillation is what OpenAI was accusing DeepSeek of doing. However, the Stanford/UW team found an ultra-low-cost way to implement it through "supervised fine-tuning." This process involves explicitly teaching the model how to reason using curated examples. Their full dataset consisted of only 1,000 carefully selected questions and solutions pulled from Google's model. TechCrunch notes that the training process took 30 minutes, using 16 Nvidia H100 GPUs. Of course, these GPUs cost a small fortune - around $25,000 per unit - but renting works out to under $50 in cloud compute credits. The researchers also discovered a neat trick to boost s1's capabilities even further. They instructed the model to "wait" before providing its final answer. This command allowed it more time to check its reasoning to arrive at slightly improved solutions. The model is not without its caveats. Since the team used Google's model as its teacher, there is the question that s1's skills, while impressive for its minuscule cost, may not be able to scale up to match the best AI has to offer just yet. There is also the potential for Google to protest. It could be waiting to see how OpenAI's case goes.

Researchers created an open rival to OpenAI's o1 'reasoning' model for under $50 | TechCrunch

AI researchers at Stanford and the University of Washington were able to train an AI "reasoning" model for under $50 in cloud compute credits, according to a new research paper released last Friday. The model known as s1 performs similarly to cutting-edge reasoning models, such as OpenAI's o1 and DeepSeek's r1, on tests measuring math and coding abilities. The s1 model is available on GitHub, along with the data and code used to train it. The team behind s1 said they created the AI model through distillation, a process to extract the "reasoning" capabilities from another AI model by training on its answers. The researchers said s1 is distilled from one of Google's reasoning models, Gemini 2.0 Flash Thinking Experimental. Distillation is the same approach Berkeley researchers used to create an AI reasoning model for around $450 last month. To some, the idea that a few researchers without millions of dollars behind them can still innovate in the AI space is exciting. But s1 raises real questions about the commoditization of AI models. Where's the moat if someone can closely replicate a multi-million dollar model with relative pocket change? Unsurprisingly, big AI labs aren't happy. OpenAI has accused DeepSeek of improperly harvesting data from its API for the purposes of model distillation. The researchers behind s1 were looking to find the simplest approach to achieve strong reasoning performance and "test-time scaling," or allowing an AI model to think more before it answers a question. These were a few of the breakthroughs in OpenAI's o1, which DeepSeek and other AI labs have tried to replicate through various techniques. The s1 paper suggests that reasoning models can be distilled with a relatively small dataset using a process called supervised fine-tuning (SFT), in which an AI model is explicitly instructed to mimic certain behaviors in a dataset. SFT tends to be cheaper than the large-scale reinforcement learning method that DeepSeek employed to train its answer to OpenAI's o1, R1. Google offers free access to Gemini 2.0 Flash Thinking Experimental, albeit with daily rate limits, via its Google AI Studio platform. Its terms forbid reverse-engineering its models to develop services that compete with Google's own AI offerings, however. We've reached out to Google for comment. S1 is based on a small, off-the-shelf AI model from Alibaba-owned Chinese AI lab Qwen, which is available to download for free. To train s1, the researchers created a dataset of just 1,000 carefully curated questions, paired with answers to those questions as well as the "thinking" process behind each answer from Google's Gemini 2.0 Flash Thinking Experimental. After training s1, which took less than 30 minutes using 16 Nvidia H100 GPUs, s1 achieved strong performance on certain AI benchmarks, according to the researchers. Niklas Muennighoff, a Stanford researcher who worked on the project, told TechCrunch he could rent the necessary compute today for about $20. The researchers used a nifty trick to get s1 to double-check its work and extend its "thinking" time: they told it to wait. Adding the word "wait" during s1's reasoning helped the model arrive at slightly more accurate answers, per the paper. In 2025, Meta, Google, and Microsoft plan to invest hundreds of billions of dollars in AI infrastructure, which will partially go toward training next-generation AI models. That level of investment may still be necessary to push the envelope of AI innovation. Distillation has shown to be a good method for cheaply recreating an AI model's capabilities, but it doesn't create new AI models vastly better than what's available today.

Researchers trained an OpenAI rival in half an hour for less than $50

To do this, researchers at Stanford 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 Verge reached out to Google with a request for comment but didn't immediately hear back.

Researchers created an AI reasoning model on par with OpenAI's o1 for less than $50

How researchers made a reasoning model on the cheap. Credit: Yuichiro Chino / Getty Images The floodgates have opened for building AI reasoning models on the cheap. Researchers at Stanford and the University of Washington have developed a model that performs comparably to OpenAI o1 and DeepSeek R1 models in math and coding -- for less than $50 of cloud compute credits. What's more, the model was trained on only 1,000 questions, and took just 26 minutes and 16 Nvidia H100 GPUs. Stanford researcher Niklas Muennighoff said in a email to Mashable that the cost is an estimate based on the GPU runtime and number of H100 GPUs used. The AI industry of late is all about how new approaches to the pre and post training process can massively save computing costs, as evidenced by DeepSeek's disruptive impact. On top of that, developers are now able to build on top of existing AI models at little or no cost, through APIs, open-source access, and even closed-source models by distilling their data, bringing the costs down even more. According to the team's research paper which was published last Friday, s1 was trained on a dataset consisting of "1,000 carefully curated questions paired with reasoning traces and answers distilled from Gemini Thinking Experimental." Google's Gemini Thinking Experimental model is accessible with daily limits through AI Studio. While it's a closed-source model, that clearly hasn't stopped researchers from making use of its responses. Next, the researchers used an "off the shelf" pretrained model from Alibaba-owned lab, Qwen, and performed supervised fine-tuning of its curated dataset. Then, the team created a token budget to control the amount of compute time for testing the model. If s1 went over budget on thinking tokens, it was cut off and forced to generate whatever answer it came up with. If the researchers wanted the model to spend more "test-time compute" on a problem, they would simply tell the model to "wait," which extended its thinking time and led to more accurate results. By controlling the amount of time and compute spent on a problem, the researchers were able to show how increased thinking team leads to improved performance. S1 is one example of open-source reasoning models that have been developed for a fraction of the cost of flagship models from Google and OpenAI. In January, UC Berkeley researchers released an open-source reasoning model called Sky-T1 that cost $450, "demonstrating that it is possible to replicate high-level reasoning capabilities affordably and efficiently," per its blog post. There's also the open-source rStar-Math reasoning model from Microsoft Asia researchers, Tulu 3 from non profit research institute Ai2, and HuggingFace has its own initiative to replicate DeepSeek's R1. As high-quality models become more accessible and cheaper, we're starting to see a power shift from the few AI heavy hitters, to the many.

New AI Reasoning Model Rivaling OpenAI Trained on Less Than $50 in Compute

It's cheap to copy already built models from their outputs, but likely still expensive to train new models that push the boundaries. It is becoming increasingly clear that AI language models are a commodity tool, as the sudden rise of open source offerings like DeepSeek show they can be hacked together on a relatively small budget. A new entrant called S1 is once again reinforcing this idea, as researchers at Stanford and the University of Washington trained the "reasoning" model using less than $50 in cloud compute credits. S1 is a direct competitor to OpenAI's o1, which is called a reasoning model because it produces answers to prompts by "thinking" through related questions that might help it check its work. For instance, if the model is asked to determine how much money it might cost to replace all Uber vehicles on the road with Waymo's fleet, it might break down the question into multiple stepsâ€"such as checking how many Ubers are on the road today, and then how much a Waymo vehicle costs to manufacture. According to TechCrunch, S1 is based on an off-the-shelf language model, which was taught to reason by studying questions and answers from a Google model, Gemini 2.0 Flashing Thinking Experimental. Google's model shows the thinking process behind each answer it returns, allowing the developers of S1 to give their model a relatively small amount of training dataâ€"1,000 carefully curated questions, along with the answersâ€"and teach it to mimic Gemini's thinking process. Another interesting detail is how the researchers were able to improve the reasoning performance of S1 using an ingeniously simple method: The researchers used a nifty trick to get s1 to double-check its work and extend its “thinking†time: They told it to wait. Adding the word “wait†during s1’s reasoning helped the model arrive at slightly more accurate answers, per the paper. This suggests that, despite worries that AI models are hitting a wall in capabilities, there remains a lot of low-hanging fruit. Some notable improvements to a branch of computer science are coming down to conjuring up the right incantation words. OpenAI has reportedly cried fowl about the Chinese DeepSeek team training off its model outputs. The irony is not lost on most people. ChatGPT and other major models were trained off data scraped from around the web without permission, an issue still being litigated in the courts as companies like the New York Times seek to protect their work from being used without compensation. Google also technically prohibits competitors like S1 from training on Gemini's outputs. Ultimately, the performance of S1 is impressive, but does not suggest that one can train a smaller model from scratch with just $50. The model essentially piggybacked off all the training of Gemini, getting a cheat sheet. A good analogy might be compression in imagery. A distilled version of an AI model might be compared to a JPEG of a photo. Good, but still lossy. And large language models still suffer from a lot of issues with accuracy, especially large-scale general models that search the entire web to produce answers. But a model like S1 could be useful in areas like on-device processing for features like Apple Intelligence. There has been a lot of debate about what the rise of cheap, open source models might mean for the technology industry writ large. Is OpenAI doomed if its models can easily be copied by anyone? Defenders of the company say that language models were always destined to be commodified. OpenAI, along with Google and others, will succeed building useful applications on top of the models. More than 300 million people use ChatGPT each week, and the product has become synonymous with chatbots and a new form of search. The interface on top of the models, like OpenAI's Operator that can navigate the web for a user, or a unique data set like xAI's access to X (formerly Twitter) data, is what will be the ultimate differentiator. Another thing to consider is that "inference" is expected to remain expensive. Inference is the actual processing of each user query submitted to a model. As AI models become cheaper and more accessible, the thinking goes, AI will infect every facet of our lives, resulting in much greater demand for computing resources, not less. And OpenAI's $500 billion server farm project will not be a waste. That is so long as all this hype around AI is not just a bubble.

This $6 AI model called s1 just challenged OpenAI's o1

A new AI model named s1, unveiled in a paper released on February 2, is garnering attention for its cost-effective performance that rivals OpenAI's o1, achieving significant capabilities at a training cost of just $6. The s1 model reaches performance levels close to state-of-the-art, utilizing simpler infrastructure. It enhances large language models (LLMs) during inference by extending "thinking time" through interventions such as replacing terminal tags with prompts like "Wait." Trained on a distilled dataset consisting of 1,000 high-quality examples from Qwen2.5, developed by Alibaba Cloud, s1 employed 16 Nvidia H100 GPUs, with a single training run lasting approximately 26 minutes. The total computational cost was about $6, allowing for more frequent experimentation, even for teams with limited resources. While larger organizations like OpenAI and Anthropic depend on extensive infrastructure, innovations like s1 demonstrate the potential for progress within constrained budgets. However, the introduction of s1 has sparked concerns regarding "distealing," a practice where models utilize distilled datasets from other AI systems, raising ethical and legal questions that have ignited industry discussions. In tests measuring math and coding abilities, s1 performs comparably to leading reasoning models, such as OpenAI's o1 and DeepSeek's R1. The s1 model, including its data and training code, is accessible on GitHub. The team behind s1 began with an off-the-shelf base model and refined it through distillation, a method to extract reasoning capabilities from an existing AI model using its answers. Specifically, s1 is distilled from Google's Gemini 2.0 Flash Thinking Experimental, representing a similar approach used by Berkeley researchers to develop an AI reasoning model for approximately $450 last month. The ability for smaller research teams to innovate in the AI space without substantial financial backing presents both excitement and challenges. The question arises about the sustainability of proprietary advantages in a landscape where costly models can be replicated affordably. OpenAI has expressed dissatisfaction, alleging that DeepSeek improperly harvested data from its API for the purpose of model distillation. Researchers aimed to devise a straightforward approach that achieves robust reasoning performance and "test-time scaling," enabling AI models to engage in deeper analysis before responding. The s1 research indicates that reasoning models can be distilled using a relatively small dataset through a method called supervised fine-tuning (SFT), which instructs the AI model to imitate certain behaviors within a dataset. This method tends to be more economical than the large-scale reinforcement learning approach utilized by DeepSeek for its R1 model. Google provides free access to Gemini 2.0 Flash Thinking Experimental, though usage is subject to daily rate limits, and its terms prohibit reverse-engineering models to create competing services. The s1 training process involved curating a dataset of 1,000 tailored questions and answers along with the reasoning processes derived from Gemini 2.0. Following training, which took less than 30 minutes, s1 demonstrated strong performance on specific AI benchmarks. Niklas Muennighoff, a researcher at Stanford involved in the project, indicated that the necessary computing resources could be rented today for about $20. The researchers also implemented a technique to enhance s1's accuracy by instructing it to "wait" during reasoning, thereby extending its thinking time and achieving slightly improved answers.

Researchers Create a Low-Cost AI Model to Analyse How OpenAI's o1 Reasons

Researchers from Stanford University and Washington University have developed an open-source artificial intelligence (AI) model that is comparable in performance to OpenAI's o1 model. The main objective of the researchers was not to create a powerful reasoning-focused model but to understand how the San Francisco-based AI firm instructed its o1 series models to perform test time scaling. Notably, the researchers were able to showcase the methodology and replicate the model's behaviour at an extremely low cost while using far fewer compute resources. The researchers detailed the methodology and process of developing the model in a study published in the pre-print journal arXiv. The process involved creating a synthetic dataset from a different AI model and using several new techniques such as ablation and supervised fine-tuning (SFT). The model is available in a GitHub listing. It should be noted that the AI model was not built from scratch. The developers used the Qwen2.5-32B-Instruct and distilled it to create the s1-32B large language model (LLM). Released in September 2024, the model is capable but given its size and lack of reasoning capabilities, it cannot match up to OpenAI's o1. During the process, the researchers used the Gemini Flash Thinking application processing interface (API) to generate reasoning traces and responses. A total of 59,000 triplets of questions, reasoning traces (the chain of thought or CoT), and responses were extracted from the API. A dataset called the s1K was then created by selecting 1,000 high-quality, diverse, and difficult questions as well as the reasoning traces and the responses. After creating the s1K dataset, the researchers performed supervised fine-tuning on the Qwen2.5-32B-Instruct model. For this, basic fine-tuning hyperparameters were used. The distillation process took 26 minutes of training on 16 Nvidia H100 GPUs. Till this point, the researchers had no idea how OpenAI trained the models to "think" and how it managed to stop the thinking process. Without this, a model runs the risk of overthinking indefinitely as it second-guesses its output wasting valuable processing power. While fine-tuning the model, the researcher found something interesting. They found that they could manipulate the inference time by adding XML tags. Once a model reaches the end tag, it is told to change its voice to an authoritative tone for the final answer. Notably, inference time is the near real-time responses that a typical AI model generates. Anything more than this would require careful manipulation of the code. With the s1-32B model, the researchers added a "wait" command to force it to think beyond the usual inference period. Once added, the model began second-guessing and verifying its output. Then, the tag was used to either shorten this test time scaling phase or lengthen it. Then, the researchers also experimented with several other phrases such as "alternatively", and "hmm", but found that the best performance metrics were achieved when using the "wait" tag. By bringing the model close to the performance of o1, the researchers claim that this might be the method used by OpenAI to fine-tune its reasoning models. A TechCrunch report claims that the researchers were able to create the s1-32B AI model under $50 (roughly Rs. 4,380), highlighting that creating a post-training structure for reasoning models can be done at an extremely low cost.

US researchers build $50 AI reasoning model, challenges OpenAI, DeepSeek

In tests involving math and coding, s1 exhibits performance comparable to cutting-edge models like OpenAI's o1 and DeepSeek's R1. "However, recent advances in reasoning, such as OpenAI's o1 and DeepSeek's r1, lack transparency, limiting broader research progress," said the research team. The researchers achieved this level of performance by employing a technique known as "distillation." This involves training s1 to replicate the reasoning abilities of another AI model, in this case, Google's Gemini 2.0 Flash Thinking Experimental model. S1 was trained on a curated dataset of 1,000 questions and answers, accompanied by the "thinking" process of the Gemini model. This allowed it to learn how to arrive at accurate solutions. "We curate a small dataset s1K of 1,000 questions paired with reasoning traces relying on three criteria we validate through ablations: difficulty, diversity, and quality," remarked the team. To optimize the training process, the researchers utilized Supervised Fine-Tuning (SFT). This method involves providing the AI model with explicit instructions and examples. This enables faster and more efficient learning compared to other techniques like Reinforcement Learning.

Academic researchers find a way to train an AI reasoning model for less than $50

A small team of AI researchers from Stanford University and the University of Washington has found a way to train an AI reasoning model for a fraction of the price paid by big corporations that produce widely known products such as ChatGPT. The group has posted a paper on the arXiv preprint server describing their efforts to inexpensively train chatbots and other AI reasoning models. Corporations such as Google and Microsoft have made clear their intentions to be leaders in the development of chatbots with ever-improving skills. These efforts are notoriously expensive and tend to involve the use of energy-intensive server farms. More recently, a Chinese company called DeepSeek released an LLM equal in capabilities to those being produced by countries in the West developed at far lower cost. That announcement sent stock prices for many tech companies into a nosedive. In this new study, the researchers claim that it is possible to train an LLM with capabilities similar to those made by OpenAI or DeepSeek for less than $50. The catch is that the researchers on this new effort used a distillation process to extract capabilities from another AI model. To train an AI so inexpensively, the research team began with an off-the-shelf AI model made by Alibaba, a China-owned company, which created the freely available test model. The research team modified the model and called the result s1. Preliminary training involved 1,000 question-and-answer pairs they had designed carefully to give their model a leg up on learning. They also gave it the "thinking process" behind Gemini 2.0, a freely available Google experimental model. They then trained it in just 26 minutes using 16 Nvidia H100 GPUs. The team also tacked on what they call a little trick -- they added a step called "thinking" that runs before the model provides an answer -- it gives the model time to double-check its work. The result, the researchers claim, is an AI model on par with other much more well-known products, made at a fraction of the cost.

Turns out, it's not that hard to do what OpenAI does for less

Even as OpenAI continues clinging to its assertion that the only path to AGI lies through massive financial and energy expenditures, independent researchers are leveraging open-source technologies to match the performance of its most powerful models -- and do so at a fraction of the price. Last Friday, a unified team from Stanford University and the University of Washington announced that they had trained a math and coding-focused large language model that performs as well as OpenAI's o1 and DeepSeek's R1 reasoning models. It cost just $50 in cloud compute credits to build. The team reportedly used an off-the-shelf base model, then distilled Google's Gemini 2.0 Flash Thinking Experimental model into it. The process of distilling AIs involves pulling the relevant information to complete a specific task from a larger AI model and transferring it to a smaller one. Recommended Videos What's more, on Tuesday, researchers from Hugging Face released a competitor to OpenAI's Deep Research and Google Gemini's (also) Deep Research tools, dubbed Open Deep Research, which they developed in just 24 hours. "While powerful LLMs are now freely available in open-source, OpenAI didn't disclose much about the agentic framework underlying Deep Research," Hugging Face wrote in its announcement post. "So we decided to embark on a 24-hour mission to reproduce their results and open-source the needed framework along the way!" It reportedly costs an estimated $20 in cloud compute credits, and would require less than 30 minutes, to train. Hugging Face's model subsequently notched a 55% accuracy on the General AI Assistants (GAIA) benchmark, which is used to test the capacities of agentic AI systems. By comparison, OpenAI's Deep Research scored between 67 - 73% accuracy, depending on the response methodologies. Granted, the 24-hour model doesn't perform quite as well as OpenAI's offering, but it also didn't take billions of dollars and the energy generation capacity of a mid-sized European nation to train. These efforts follow news from January that a team out of University of California, Berkeley's Sky Computing Lab managed to train their Sky T1 reasoning model for around $450 in cloud compute credits. The team's Sky-T1-32B-Preview model proved the equal of early o1-preview reasoning model release. As more of these open-source competitors to OpenAI's industry dominance emerge, their mere existence calls into question whether the company's plan of spending half a trillion dollars to build AI data centers and energy production facilities is really the answer.