DeepSeek R1 open source reasoning model challenges OpenAI dominance

A new class of artificial intelligence models that 'think' before they answer is dismantling the competitive moat built by Silicon Valley's richest labs. The release of DeepSeek-R1, an open-source reasoning model from a Chinese research firm, has demonstrated that the advanced logical capabilities pioneered by OpenAI can be replicated, and distributed freely, at a fraction of the training cost.

The development marks a major shift in how foundation models are built and run. For years, the AI industry operated on the belief that better performance required ever-larger datasets and increasingly massive supercomputers. DeepSeek-R1, alongside OpenAI's proprietary o1 and o3 series, proves that a different approach—known as test-time compute—can yield massive jumps in capability without a corresponding explosion in raw computing power.

How reasoning models work

Traditional large language models are essentially highly sophisticated predictive text engines. They generate the next most likely word in a fraction of a second, without any ability to pause, plan, or correct course mid-sentence.

Reasoning models change this by using a system of reinforcement learning to generate an internal 'chain of thought' before presenting a final answer. When faced with a complex coding or mathematical problem, the model drafts intermediate steps, tests its own logic, identifies errors, and refines its approach. The user only sees the final output, or a collapsed dropdown of the model's internal monologue.

This hidden deliberation allows the models to solve multi-step problems that routinely trip up standard chatbots. On standard academic benchmarks for mathematics and computer programming, these reasoning models are performing at levels comparable to PhD students.

The economic disruption of open source

While OpenAI was the first to commercialise this capability with its o1 model, DeepSeek's decision to open-source the weights of R1 under a permissive MIT licence has transformed the market dynamics.

Building and training frontier models typically costs hundreds of millions of dollars. DeepSeek claimed to have trained its model for a fraction of that figure—reportedly under $6 million—by using innovative training architectures and relying heavily on reinforcement learning rather than brute-force pre-training.

By making the model open-source, the firm has effectively commoditised reasoning. Independent developers and rival tech firms can now run, modify, and build upon these reasoning capabilities without paying subscription fees or API tolls to a single gatekeeper.

What happens next

The rapid adoption of reasoning models is already forcing a reassessment of hardware demands and infrastructure investment.

Because these models spend more time 'thinking' during inference, the demand for computing power is shifting from the training phase to the run-time phase. This has led to a flurry of development in inference optimisation, with researchers finding ways to run distilled, smaller versions of these reasoning models directly on consumer laptops and smartphones.

However, significant hurdles remain. Reasoning models are noticeably slower than traditional LLMs, making them less suitable for simple conversational tasks. They also suffer from 'over-thinking,' sometimes generating thousands of words of internal reasoning for straightforward queries, which drives up operational costs.

The industry has not yet resolved whether this reasoning process genuinely understands logic, or if it is simply mimicking the structure of logical thinking at a grander scale. What is clear, however, is that the race to build the next generation of AI is no longer just about who has the biggest cluster of microchips.