Stanford’s $50 AI Model Questions the Multi-Billion Dollar AI Arms Race

Stanford researchers have developed an AI model for just $50 that rivals OpenAI and Google’s top reasoning models, challenging the high-cost AI race.

By Markus Kasanmascheff

February 6, 2025 9:56 am CET

For years, artificial intelligence research has been dominated by companies pouring billions into massive AI models, assuming that sheer computational power would keep them ahead. But a new project from Stanford University and the University of Washington is challenging that belief.

Their latest model, s1, was trained for less than $50 in compute costs, yet performs competitively with reasoning AI models developed by OpenAI and DeepSeek.

Unlike proprietary models that require extensive infrastructure and months of training, s1 was fine-tuned in under 30 minutes using just 16 Nvidia H100 GPUs, according to the researchers.

Its code, methodology, and dataset have been made available through an open-source GitHub repository, making it accessible for anyone to inspect, replicate, or improve. The project raises a critical question for the AI industry: Is a multi-billion-dollar budget still necessary to compete at the highest level?

A Model That Puts OpenAI and Google’s AI Strategies at Risk

AI giants like OpenAI, Google, and Microsoft have bet heavily on their ability to outspend competitors in AI model training and infrastructure.

OpenAI’s o1 model and Google’s Gemini 2.0 Flash are designed with this advantage in mind. However, s1’s development proves that high-level reasoning capabilities can be replicated at a fraction of the cost.

The research team behind s1 used a technique called distillation, where a smaller model is trained to mimic the responses of a larger AI system.

Instead of developing an AI model from scratch, they took Qwen2.5-32B-Instruct, a freely available model from Alibaba’s Qwen AI Lab, and fine-tuned it using 1,000 carefully selected math and reasoning questions.

Notably, the dataset was generated using Google’s Gemini 2.0 Flash Thinking Experimental model. As stated in the s1 research paper, “we construct s1K, which consists of 1,000 carefully curated questions paired with reasoning traces and answers distilled from Gemini Thinking Experimental.”

While Google provides free API access to this model, its terms of service prohibit using its outputs to develop competing AI models. The company has not yet commented on whether s1 violates these restrictions.

Performance That Matches or Surpasses Commercial Models

Despite being trained on a relatively small dataset, s1 achieves performance levels comparable to OpenAI and DeepSeek’s models.

On the AIME24 benchmark, which measures AI math problem-solving ability, s1 achieved a 56.7% accuracy score, outperforming OpenAI’s o1-preview, which scored 44.6%. Similarly, on the MATH500 benchmark, s1 reached 93% accuracy, matching the results of DeepSeek R1.

However, the model shows some limitations in broader scientific knowledge. On the GPQA-Diamond benchmark, which contains advanced physics, biology, and chemistry problems, s1 scored 59.6%, falling behind OpenAI and Google’s models.

s1 model performance benchmarks compared to leading AI models from Google and OpenAI

Still, for a model trained in under 30 minutes with minimal compute, these results challenge the assumption that bigger datasets and longer training cycles are always necessary.

An Unexpected Trick That Improves AI Reasoning

The research also revealed an unexpected finding that improved s1’s accuracy. Instead of modifying the model itself, they experimented with the way prompts were structured.

The study states, “we develop budget forcing to control test-time compute by forcefully terminating the model’s thinking process or lengthening it by appending ‘Wait’ multiple times to the model’s generation when it tries to end. This can lead the model to double-check its answer, often fixing incorrect reasoning steps.”

Simply adding the word “wait” into the prompt forced s1 to spend more time considering its response before finalizing an answer. This approach aligns with recent research into test-time scaling, where models improve accuracy by allocating more computation to complex tasks instead of responding instantly.

Could OpenAI and Google Start Locking Down Their AI Models?

The rise of low-cost AI reasoning models like s1 presents a challenge for companies that have invested heavily in exclusive AI systems.

OpenAI and Google have argued that building reliable and safe AI models requires significant compute resources, justifying their premium AI services and restrictive access policies.

However, as more researchers demonstrate that high-level AI capabilities can be replicated cheaply, these companies may look for new ways to protect their models from being reverse-engineered or distilled into smaller competitors.

OpenAI has already shown signs of tightening access to its technology. The company currently restricts its Deep Research feature to paid ChatGPT Pro users, limiting external AI developers’ ability to study its methods. Google, meanwhile, imposes strict rate limits on access to its Gemini 2.0 API and explicitly forbids training competing AI models using its outputs.

With projects like s1 emerging, there is a growing likelihood that companies will implement watermarking techniques or legal restrictions to prevent their AI-generated outputs from being used for training other systems. However, enforcing these rules in open-source AI research environments will be extremely difficult.

The Future of AI: Open Research or Corporate Control?

As AI research continues to advance, the battle between open-source innovation and proprietary AI development is becoming more intense. The success of distilled AI models like s1 and Sky-T1 suggests that AI capabilities are no longer exclusive to tech giants.

Major AI companies argue that proprietary models provide better control over AI risks, ensuring safety, bias reduction, and regulatory compliance. But independent researchers counter that open-source models improve transparency, allowing experts to audit and refine AI systems without corporate influence.

Governments and regulators are also closely watching these developments. AI policymaking has so far focused on governing large-scale models, but the emergence of low-cost AI replication techniques could shift the conversation toward data access restrictions and ethical considerations.

The release of s1 signals a shift that could reshape the AI industry. If powerful reasoning AI can be replicated for under $50, smaller AI research teams and startups may soon have the ability to compete with billion-dollar AI companies.

For now, s1 remains open-source, meaning researchers worldwide can test, modify, and expand upon its capabilities. However, if OpenAI, Google, and other AI labs see this as a threat, they may push for stricter API access controls, licensing restrictions, or even legal action against AI distillation methods.

Will the future be defined by corporate-controlled, proprietary models, or will open AI research continue to advance, making high-level AI reasoning accessible to all? Let us know in the comments what you think.