DeepSeek R1 reasoning model challenges OpenAI dominance at a fraction of the cost

The competitive balance in artificial intelligence has shifted. DeepSeek, a low-profile research firm based in Hangzhou, has released DeepSeek-R1, a reasoning model that matches the performance of OpenAI's flagship o1 on key benchmarks while costing only a fraction of the price to run.

Unlike traditional large language models that generate immediate, word-by-word answers, reasoning models use a 'chain of thought' to think through complex problems before responding. DeepSeek-R1 uses reinforcement learning to develop these reasoning patterns natively. The result is a model that can tackle complex mathematics, coding, and scientific logic at a level previously reserved for the most expensive proprietary systems.

How DeepSeek-R1 achieves reasoning

Most foundation models rely on massive, expensive datasets of human-labelled demonstrations to learn how to behave. DeepSeek's researchers bypassed much of this costly stage. They allowed their model to explore, make mistakes, and self-correct through pure reinforcement learning, rewarding it for reaching the correct logical conclusion.

During training, the model naturally developed reasoning behaviours. It learned to double-check its work, re-evaluate initial assumptions, and even translate thoughts into different languages mid-process to help it solve problems. This development suggests that high-end AI reasoning does not require the multi-billion-dollar budgets of Silicon Valley's tech giants.

The cost comparison

The most disruptive aspect of DeepSeek-R1 is its price. DeepSeek is offering access to its model through an API at roughly $0.55 per million input tokens and $2.19 per million output tokens. For context, developers using OpenAI's o1 pay significantly more—often up to $15 per million input tokens and $60 per million output tokens.

While these prices will inevitably shift as competition intensifies, the initial margin is stark. It suggests that the cost of advanced inference—the process of running a trained model—is falling much faster than the industry anticipated.

Open-source distillation

In a move that has surprised the open-source community, DeepSeek has released the model weights for R1 under an open-source MIT licence. The firm has also published 'distilled' versions of the model, ranging from 1.5 billion to 70 billion parameters, trained using R1's outputs.

These smaller models, built on foundations from Meta's Llama and Alibaba's Qwen, allow researchers to run advanced reasoning capabilities on consumer-grade hardware. It means developers do not need access to massive server farms to build applications that can reason.

Limitations and what happens next

Despite the breakthrough, significant hurdles remain. DeepSeek-R1 is highly capable at logic and math, but users have noted that it can struggle with creative writing, nuanced conversation, and language tasks that require stylistic sensitivity. There are also unresolved questions about safety filters, as well as the geopolitical tensions surrounding a Chinese firm matching American AI leaders.

Yet the release of R1 marks a clear departure from the status quo. It proves that reasoning capabilities are no longer the exclusive domain of proprietary API providers, forcing the entire AI market to rethink its pricing and development strategies.