DeepSeek-R1: reasoning as a new dimension of competition

In January 2025, Chinese company DeepSeek released the R1 model and made its weights open. This is not a routine research event. It is a moment after which the conversation about the cost and accessibility of advanced AI capabilities shifts sharply.

I have seen a few such shifts in technology before. The emergence of R1 resembles the moment when cloud providers began offering compute at prices that made owning a data center indefensible for most companies. The difference is that this cycle compressed into a matter of months.

What reasoning models are and why it matters

Until recently, large language models worked on one principle: receive a prompt, produce an answer. Fine for text generation, translation, summarisation. Poor for tasks requiring step-by-step analysis: mathematics, logic, complex analytical conclusions.

Reasoning models are a different class. Before answering, they expand a chain of intermediate steps, check themselves, revise the approach. The result holds together better on hard problems. This is what OpenAI o1 does, and R1 now covers the same class of tasks.

For business, what matters is not how this works internally. What matters is that tasks which a year ago required an expensive commercial model with API access can now be handled with an open model deployed locally or at any provider for significantly less money.

What changes in the economics of AI projects

Before R1, the typical conversation about deploying AI for analytical tasks ran into the cost of API calls to commercial models. At high request volumes, that cost became an argument against starting a pilot or against scaling.

The situation is different now. Open weights mean:

• the model can be deployed independently, without depending on a single provider;
• inference pricing falls due to competition among hosting providers;
• there is no lock-in to one company's pricing policy and terms of service;
• for tasks involving sensitive data, it becomes possible to work without sending data outside the company's perimeter.

This does not mean commercial models disappear. They remain stronger for certain tasks and more convenient for a fast start. But monopoly pricing on reasoning capabilities is over.

What this means for companies that have not started yet

If a company is still watching from the sidelines - looking at AI for analytics, documentation, request handling, or assistants - the moment for a pilot has become significantly cheaper.

Practical points worth considering when planning:

First - when estimating the cost of an AI project, there is no longer any reason to base it on a single provider's pricing. The market is competitive now, and it is worth measuring before making a decision.

Second - open weights change the conversation about data confidentiality. For tasks where data cannot leave the company's perimeter, this opens up possibilities that barely existed a year ago.

Third - the pace of change in the model market means that decisions about AI infrastructure should be made with a short review horizon. What was the expensive default at the start of 2024 is neither the default nor expensive by mid-2025.

What has not changed

A lower model cost does not solve the data problem. More accessible reasoning capabilities do not automatically make projects more useful when the underlying data is fragmented, poorly structured, or missing.

A cheaper model on bad data produces cheaper but equally unreliable results.

The right question for a manager right now is not "should we start an AI project given model costs". Model costs are no longer the main barrier. The main question is how ready the data and processes that will feed the model actually are.