DeepSeek-R1 challenges OpenAI with low-cost open-source reasoning model

The Chinese research firm DeepSeek has upended the artificial intelligence industry by releasing DeepSeek-R1, an open-source reasoning model that performs on par with OpenAI’s proprietary o1. Released under the permissive MIT licence, the model offers the developer community free access to the kind of advanced, "slow-thinking" reasoning capabilities that were previously locked behind expensive corporate APIs.

The release has triggered intense debate in Silicon Valley about the economics of AI development. For the past two years, the prevailing belief was that building leading-edge frontier models required billions of dollars, tens of thousands of Nvidia GPUs, and massive power grids. DeepSeek has challenged that assumption by proving that highly efficient training techniques can yield world-class reasoning capabilities at a fraction of the cost.

How does DeepSeek-R1 work?

Unlike traditional large language models that generate the next most likely word instantly, DeepSeek-R1 uses reinforcement learning to "think" before it writes. When presented with a complex prompt, it generates an internal chain of thought—testing hypotheses, correcting its own mistakes, and refining its logic—before delivering its final answer.

This reasoning behaviour makes it particularly potent for mathematics, computer programming, and scientific logic. The model is released alongside distilled versions, which port its reasoning capabilities into smaller, more efficient models developed by Meta and Alibaba, allowing them to run on consumer-grade hardware.

Why the training cost matters

The most disruptive aspect of DeepSeek's breakthrough is the financial efficiency of its development. While US tech giants have not disclosed the exact training costs of their latest frontier models, industry analysts estimate they run into the hundreds of millions of dollars. DeepSeek claims to have trained its base model, DeepSeek-V3, for just $5.6 million, using reinforcement learning to build R1 on top of it.

If these figures are accurate, it suggests that algorithmic efficiency and smart training techniques can bypass the brute-force computational scaling that has dominated AI development. It also raises questions about the long-term competitive moat of companies whose business models rely on charging high premiums for closed-source inference.

What are the performance benchmarks?

According to technical reports published by the company, DeepSeek-R1 matches or exceeds OpenAI's o1 across several major benchmarks. On the American Invitational Mathematics Examination (AIME), a standard test for high-school math talent, R1 achieved a score of 79.8 per cent, compared to o1's reported 74.4 per cent. On Codeforces, a competitive programming platform, it achieved a rating that places it in the 96th percentile of human participants.

However, independent verification of these claims is still ongoing. Early testing by developers suggests that while R1 is exceptionally strong at logical tasks, it can struggle with creative writing, stylistic consistency, and certain safety guardrails. There are also unresolved questions about how the model behaves when faced with complex multi-turn conversations outside of scientific and mathematical domains.

What happens next?

The arrival of DeepSeek-R1 is likely to accelerate the commoditisation of reasoning capabilities. With the weights of the model freely available, developers are already integrating R1 into autonomous AI agents, local coding assistants, and specialised research tools without the risk of vendor lock-in.

It also carries geopolitical implications. DeepSeek's ability to train a frontier-class model on older, less powerful hardware suggests that US export controls on high-end chips may not be the insurmountable barrier to Chinese AI progress that policymakers had hoped.