DeepSeek and OpenAI spur race towards reasoning models as test-time compute takes over

Frontier AI development has quietly changed direction. For years, progress relied on scaling up pre-training: gathering larger datasets, assembling bigger GPU clusters, and letting models predict the next token across broader swathes of the internet. That approach has run into diminishing returns, power constraints, and a shortage of fresh human text.

Instead, labs are now spending their computing power when the model runs, not just when it trains. The release of reasoning systems such as OpenAI's o1 and o3, alongside the rapid open-weight release of DeepSeek-R1 from China, has made chain-of-thought processing the central battleground for artificial intelligence.

What changes with reasoning models

Traditional language models generate an answer immediately, predicting one word after another without internal revision. If a calculation or a line of code goes wrong early in the output, the model tends to hallucinate its way to a conclusion rather than correct itself.

Reasoning models work differently. They spend compute during inference to generate hidden intermediate tokens—often called a chain of thought—before showing an answer to the user. They test hypotheses, back up when an approach fails, and correct errors in real time.

On benchmark tests that measure genuine problem-solving rather than rote memorisation, the difference is stark. In competitive programming evaluations such as Codeforces and mathematical competitions like the American Invitational Mathematics Examination (AIME), models using inference-time compute score substantially higher than conventional models of identical parameter size.

How DeepSeek-R1 shifted the competitive balance

Until January, advanced chain-of-thought techniques were largely confined to proprietary systems behind commercial APIs. DeepSeek changed that by releasing the weights and training recipes for DeepSeek-R1.

The Chinese research team demonstrated that large-scale reinforcement learning alone, applied directly to base models without requiring massive volumes of human-curated chain-of-thought demonstrations, could teach a system to reason out loud. DeepSeek also released distilled versions down to small architectures such as 1.5 billion and 8 billion parameters, allowing individual researchers and smaller companies to run competitive reasoning tools locally on consumer hardware.

The release has triggered sharp debate over the costs of frontier research. While leading US firms spent billions building proprietary clusters, DeepSeek reported spending less than six million dollars in direct computing costs to train its underlying V3 foundation model, though independent analysts have pointed out that total infrastructure outlays across the company remain difficult to verify independently.

Where frontier labs go next

Competitors are already reacting to the shift. Anthropic, Google, and Meta have accelerated plans to introduce explicit test-time thinking into their flagship architectures.

The next step is extending reasoning beyond pure text and math into multimodal tasks and agent workflows. Researchers are attempting to pair inference-time search with tool use, enabling models to browse software environments, run code in sandboxes, and inspect error logs before returning an answer.

That transition also creates commercial friction. Reasoning tokens cost money to generate, even if the user never sees them. As inference costs climb with every difficult query, the business model of flat-rate subscriptions is coming under strain, forcing providers to reconsider how they price access to their most capable models.