DeepSeek R1 reasoning model challenges OpenAI dominance in high-end AI

The artificial intelligence market has been upended by the arrival of high-performing, open-source reasoning models that match the capabilities of the world's most expensive proprietary systems. At the centre of this shift is DeepSeek, a Hangzhou-based research company, which released its R1 model. The release has forced a reassessment of the competitive dynamics between closed-source pioneers like OpenAI and the broader open-source community.

What is the DeepSeek R1 breakthrough?

DeepSeek R1 is a reasoning model. Unlike traditional large language models that predict the next word in a fraction of a second, reasoning models use "test-time compute". They pause, generate an internal chain of thought, correct their own mistakes, and evaluate different approaches before delivering an answer.

In benchmark tests published by the company, DeepSeek R1 performed on par with OpenAI’s o1 model across mathematics, coding, and scientific reasoning tasks. On the MATH-500 benchmark, DeepSeek reported that R1 achieved a 97.3% accuracy rate, compared to OpenAI o1’s 94.8%. While these benchmarks are self-reported and have not been independently audited, third-party developers replicating the results have confirmed a substantial leap in capability over previous open-source systems.

Why the cost of training matters

The most disruptive aspect of the release is not just what the model can do, but how much it cost to build. DeepSeek claims to have trained its underlying 671-billion-parameter mixture-of-experts model for approximately $5.6 million. In contrast, training runs for comparable frontier models in the US are widely estimated by industry analysts to cost hundreds of millions of dollars.

DeepSeek achieved this efficiency through a combination of novel training techniques. It used multi-head latent attention to reduce memory usage during inference and implemented a training pipeline that bypassed the need for massive, expensive supervised fine-tuning datasets. Instead, the model learned how to reason primarily through pure reinforcement learning, discovering logical patterns on its own.

The squeeze on proprietary AI models

This dramatic reduction in cost challenges the prevailing belief that building leading-edge AI requires infinite capital and massive compute clusters. If a relatively small Chinese firm can match Silicon Valley’s best reasoning models on a shoestring budget, the competitive "moat" around proprietary software giants looks increasingly fragile.

Developers are already using distilled versions of R1—smaller, highly efficient models that run on consumer-grade hardware but inherit the reasoning patterns of the larger model. This democratises access to advanced reasoning, allowing companies to build local, private AI systems without paying subscription fees or API costs to US tech firms.

What happens next for AI agents

The availability of cheap, open-source reasoning is expected to accelerate the deployment of autonomous AI agents. True agents require the ability to plan, use tools, and recover from errors. Traditional models were too unreliable and expensive to be left running in the background to solve complex, multi-step problems.

By lowering the cost of reasoning by orders of magnitude, models like R1 make agentic workflows commercially viable. Software developers are already experimenting with agents that can autonomously debug large codebases or conduct market research over several hours, using reasoning models to guide their actions. The race has now shifted from raw model size to the practical application of cheap intelligence.