DeepSeek R1 reasoning model challenges OpenAI dominance in AI arms race

The assumption that frontier AI development requires endless billions of dollars and exclusive access to the world's largest supercomputers has been shattered.

China's DeepSeek has released R1, a reasoning model that matches OpenAI’s proprietary o1 across several key benchmarks, publishing the weights openly for anyone to download and run.

The development marks a major shift in how the industry builds and runs artificial intelligence. Instead of simply predicting the next word based on sheer size—the brute-force scaling laws that dominated the industry for years—DeepSeek R1 and OpenAI o1 rely on "test-time compute". They are trained to pause, generate an internal chain of thought, and correct their own mistakes before delivering an answer.

Why reasoning models change the game

Traditional large language models struggle with multi-step logic, often hallucinating when asked to solve complex math or coding problems. Reasoning models bypass this by using reinforcement learning to reward the model for logical thinking patterns during training.

The results are stark. On the MATH-500 benchmark, DeepSeek R1 scored an impressive 97.3 per cent, placing it neck-and-neck with OpenAI’s o1. On competitive programming tasks like Codeforces, it reached the 96.3rd percentile.

But the real shockwave is the price tag. DeepSeek claims the training cost for its base model, DeepSeek-V3, was roughly $5.6 million. While the company has not disclosed the exact cost of training the R1 reasoning layer on top of that, industry analysts estimate it is a tiny fraction of the hundreds of millions of dollars Microsoft and OpenAI are believed to have spent on their latest systems.

The open-source disruption

By releasing R1 under an MIT licence, DeepSeek has effectively democratised reasoning capabilities. Developers can now inspect the model's thinking process directly—something OpenAI keeps strictly hidden behind its API.

The move has forced US tech giants to defend their proprietary moats. It is not yet clear how OpenAI or Google will respond to the pricing pressure, but the cost of running reasoning queries has already started to fall. Startups are already using distilled versions of R1—smaller models trained on its outputs—to run advanced reasoning tasks on standard consumer hardware.

What happens next

The success of R1 suggests the competitive moat in AI may not be computing power alone, but how efficiently that power is used. Attention is now turning to agentic workflows.

With reasoning models acting as the core brain, AI agents can now reliably plan, write code, run tests, and debug their own work without human intervention. The next battleground will not just be about scoring points on math tests, but about which model can autonomously manage complex corporate workflows for hours at a time. For now, the race is no longer a private Silicon Valley club.