DeepSeek-R1 open-source model challenges OpenAI dominance in reasoning AI

DeepSeek, a Chinese artificial intelligence company, has released an open-source model called DeepSeek-R1 that matches the reasoning capabilities of its most advanced US rivals at a fraction of the training cost.

For the past year, the AI industry has operated on the assumption that building frontier-class models requires tens of thousands of specialised chips, gigawatts of power, and billions of dollars in capital. DeepSeek-R1 has upended that narrative. By matching the performance of proprietary American systems for a tiny fraction of the expense, the release has forced a fundamental rethink of the technology's economic reality.

What is DeepSeek-R1?

Like OpenAI's o1, the new model uses a technique known as "inference-time compute" or reasoning. Instead of spitting out an answer instantly, the model generates a visible "chain of thought." It literally thinks out loud on the screen, testing hypotheses, catching its own mistakes, and refining its logic before presenting a final response to the user.

This approach makes the model exceptionally good at tasks that require structured logic, such as computer programming, advanced mathematics, and scientific problem-solving. On standard academic tests, the model performs at the very top tier. DeepSeek claims R1 achieved a 97.3 per cent score on the MATH benchmark, putting it on par with OpenAI's o1.

The shift in training economics

The most significant aspect of the release is not just what the model can do, but how it was built. While Silicon Valley firms are planning data centres costing tens of billions of dollars, DeepSeek reportedly trained its underlying model for roughly $5.6 million.

To achieve this, the company relied heavily on reinforcement learning—a process where the AI learns through trial and error, guided by rewards for correct answers—rather than relying solely on massive, expensive datasets curated by humans. This algorithmic efficiency allowed DeepSeek to squeeze maximum performance out of older, less powerful hardware, bypassing some of the physical bottlenecks that US export controls were intended to impose on Chinese tech firms.

What happens next?

By releasing the model under a permissive MIT licence, DeepSeek has effectively democratised reasoning-class AI. Developers, academic researchers, and rival businesses can now download, run, and modify the model for their own needs without paying ongoing API fees to a US provider.

It is not yet clear how American laboratories will respond. While some industry analysts suggest this breakthrough will force proprietary giants like Google and Anthropic to lower their prices, others believe it will trigger an even more aggressive push toward massive computing clusters to maintain a raw performance lead. What is certain is that the barrier to entry for world-class reasoning models has just been permanently lowered.