OpenAI vs DeepSeek. Who is stronger?

A Chinese startup has unveiled an AI model, R1, that rivals OpenAI’s technology, but with significant differences.

DeepSeek positions its latest AI model R1 as particularly effective for solving complex tasks, comparable in capability to OpenAI’s o1 reasoning model, while operating at significantly lower costs per request.

It’s impossible to consider DeepSeek’s new AI model without comparing it to OpenAI, its primary American competitor.

DeepSeek claims that their R1 model excels at answering complex questions, demonstrating a performance level on par with OpenAI’s o1, but operating at a fraction of the cost. Currently, DeepSeek's app ranks first in downloads on the US App Store.

On Wednesday, OpenAI reported that it is investigating whether DeepSeek trained its chatbot by repeatedly querying OpenAI’s models.

Just as DeepSeek and OpenAI are fundamentally different companies, their R1 and o1 models employ distinct technologies. Here are five key similarities and differences between them.

How DeepSeek Works

DeepSeek has reduced the amount of data processing required to train its models by applying both proprietary developments and techniques adapted from other Chinese AI companies operating under similar constraints, according to The Wall Street Journal.

In addition to reducing data volumes—which significantly saves time and computational power—DeepSeek employs a "Mixture of Experts" (MoE) approach. This method, also used by other AI developers, routes different requests to specialized “experts” within the model. Each expert requires less training, thus reducing the burden on hardware and improving system efficiency.

"The methods they applied aren't new, but implementing them on such a scale and with such confidence is genuinely innovative," said Luke Arrigoni, CEO of Loti AI, a company specializing in AI solutions for internet privacy.

This approach reduces computational resources during request processing but increases resource consumption when generating responses. DeepSeek uses the "Chain-of-Thought Reasoning" (CoT) method, which allows the model to solve complex tasks step-by-step, explained Lin Qiao, CEO and co-founder of Fireworks AI.

OpenAI’s o1 model also uses this method but does not reveal its reasoning process to users, Qiao added. She noted that DeepSeek’s key differentiator is that its model not only displays its reasoning process but can also use these insights to train more compact AI models.

Both models, DeepSeek's R1 and OpenAI’s o1, are capable of performing reasoning tasks, such as writing business plans or creating crossword puzzles.

Performance

DeepSeek researchers claim that they benchmarked R1 against OpenAI's leading AI models and found it to be highly competitive. One of the tests included a method developed by OpenAI, where AI models had to complete programming tasks, such as debugging code.

R1 demonstrated performance comparable to OpenAI's o1 and outperformed an earlier version, o1-mini.

According to Lin Qiao, open-source community members have already created a lightweight version of R1 that can run on mobile phones and tablets.

Some users noted that R1’s writing and problem-solving skills are impressive but that it lags behind OpenAI's o1 in specific tasks.

On Monday, OpenAI CEO Sam Altman called R1 "an impressive model, especially given its cost-efficiency," in a post on X (formerly Twitter). He also stated that the emergence of such a competitor motivates OpenAI to accelerate the release of new products.

Cost

DeepSeek claims to have achieved results comparable to OpenAI’s models at significantly lower costs without using top-tier chips.

According to some estimates, training one of DeepSeek’s earlier models required only about $5 million for chip procurement. However, Bernstein Research analyst Stacy Rasgon pointed out in a report that these estimates do not account for the costs of research and experimentation necessary for model development.

Exact details on the computational resources used to train R1 are still unknown.

By comparison, OpenAI stated that training GPT-4 cost over $100 million, and future AI models could exceed $1 billion.

Estimates suggest that training OpenAI’s next model, GPT-5, might take about six months and cost around $500 million in computational resources alone.

Privacy and Security

Users of DeepSeek’s latest flagship model, V3, noticed that it refuses to answer politically sensitive questions about China and its leader Xi Jinping. In some cases, its responses align with Beijing’s official rhetoric, whereas ChatGPT offers alternative viewpoints, including those of government critics.

Nevertheless, the R1 model is available for free download and use, leading some users to deploy it on their own servers or on servers based in the United States. Luke Kim, CEO of startup Liner, stated that his company is considering using R1 since it is open-source and can be easily replaced by other AI models.

In contrast, OpenAI reported that it has developed a "new safety training approach" to ensure that its o1 model adheres to corporate standards.

The company also emphasized its efforts to prevent AI security breaches by signing official agreements with AI safety institutions in the US and UK.

"Jailbreaking" AI involves manipulating or bypassing a model's security measures.

Open Source vs. Proprietary Solutions

DeepSeek has released the "weights" (numerical parameters) of its R1 model, allowing users to freely download, use, and modify it. However, the company has not published the dataset used for training, leading some experts to argue that the model cannot be considered fully open-source.

The Chinese company also published a report on its model training process, which AI specialists say helps developers understand DeepSeek’s innovative solutions.

Releasing model weights means developers can load and use the model in their projects. Hugging Face, the largest open-source AI model platform, reported that community-built R1 models have been downloaded 3.2 million times.

In contrast, OpenAI’s o1 model is proprietary. This means that users and companies must pay for access to the model and its capabilities.

Some companies prefer proprietary technologies because they undergo creator verification and include built-in cybersecurity mechanisms. Others choose open-source solutions due to their easier customization and adaptability to specific needs.

Companies that prioritize flexibility often opt for open-source models like DeepSeek’s R1 because they can tailor the AI to meet their unique requirements. However, proprietary models like OpenAI's o1 offer pre-integrated security and support features, making them more appealing to enterprises focused on compliance and data protection.

This divide has sparked discussions among tech leaders regarding the future of AI development. While DeepSeek has gained traction with open-source communities, OpenAI's proprietary approach continues to dominate corporate sectors that demand stringent safety and privacy controls.

The competition between the two companies is likely to shape the AI landscape in the coming years. As startups like DeepSeek challenge established players, the push for innovation, cost reduction, and customization will intensify, giving users more options tailored to different use cases and budgets.