DeepSeek-R1 and OpenAI o1 shift the AI race from larger models to longer thinking

The competitive arena of artificial intelligence has undergone its most significant shift since the debut of ChatGPT. The race is no longer primarily about feeding larger volumes of internet data into massive neural networks during the initial training phase. Instead, the industry has pivoted to "reasoning" models—systems designed to pause, generate internal chains of thought, and correct their own mistakes before presenting an answer.

This transition, pioneered by OpenAI's o1 and o3 models, was accelerated by the release of DeepSeek-R1, an open-source reasoning model developed by a Hangzhou-based startup. The launch has forced a reassessment of AI development. DeepSeek-R1 not only matches OpenAI's proprietary systems on advanced maths and coding benchmarks, but its creators claim to have trained it for a fraction of the cost of typical Silicon Valley frontier models.

Why reasoning models matter

Previously, a large language model's capabilities were largely fixed once its training was complete. To build a smarter model, companies had to spend hundreds of millions of dollars on massive GPU clusters, energy, and data.

Reasoning models change this calculation through a concept known as test-time compute, or inference scaling. Instead of generating the very first word that comes to mind, these models use reinforcement learning to explore different reasoning paths. They draft a solution, check it for logical errors, and revise their approach.

By spending more processing power at the moment of query—essentially "thinking" for several seconds or even minutes—a smaller model can achieve results that previously required far larger, more expensive systems. This allows developers to trade time for intelligence.

Benchmark performance and the open-source threat

On standard mathematics and programming benchmarks, reasoning models have shown a steep trajectory of improvement. On the American Invitational Mathematics Examination (AIME), a contest designed for top high school students, DeepSeek-R1 reportedly solved 97.3% of the problems, putting it on par with OpenAI’s o1 model. On coding platforms such as Codeforces, these systems regularly place in the top percentiles of human participants.

The fact that DeepSeek has made its model weights freely available to download has disrupted the proprietary business models of Western technology companies. If high-tier reasoning can be run locally or via cheap open-source alternatives, the premium prices charged by proprietary API providers may prove difficult to sustain.

The path to AI agents

This shift in architecture is expected to accelerate the development of autonomous AI agents. Traditional language models are poorly suited for multi-step tasks, such as booking a flight or debugging a large software repository, because a single error early in the process causes the entire run to fail.

Reasoning models, by contrast, can plan ahead and verify their own work. If a piece of code they write fails a test, they can read the error message and rewrite the code before showing it to the user.

However, these capabilities introduce new trade-offs. Reasoning models are noticeably slower to respond and require significantly more computing power during everyday use than standard LLMs. For simple tasks, like drafting an email or summarising a document, this extra processing is both unnecessary and costly. The next challenge for researchers is to build systems that can automatically decide when a question requires deep, slow thought, and when a quick, cheap response will do.