The race in artificial intelligence has long revolved around one question: who has the most capable models. Each new generation seeks to solve more complex tasks, while comparisons continue to look at the top of the rankings. But there is another question that starts to gain importance when we stop trying these tools out of curiosity and actually need to use them: How much does it cost us to get a good enough answer?. The difference may seem small, until we imagine the same request being repeated thousands or millions of times.
There are more paths to this choice than those offered by the artificial intelligence applications we use every day. We can access these tools through free services, paid subscriptions or providers that charge per use, but there are also open source models that can be run on our own infrastructure or through third parties. For a company, this opens the door to possibilities that go beyond employing the most advanced services on the market.
The discussion led to the presentation of IBM’s financial results. During question period, Matt Swanson, an analyst at RBC Capital Markets, asked how big companies are balancing two demands that could come into conflict: spending on artificial intelligence and maintaining control over its deployment. Arvind Krishna, CEO of IBM, responded by going back to the end of 2024 and comparing the evolution of the most advanced models. Open weight option. His argument points to a shift that, he believes, is changing that choice.
There’s a new battle in artificial intelligence, and it’s not just fought in performance
To understand the magnitude of the change the executive described, it’s worth looking at the numbers he used. At the end of 2024, he raised an example where a frontier model could receive 70 points in a comprehensive assessment, compared to 20 or 30 for an open option. “It’s a huge gap. No one will accept it,” he said. By the end of 2025, according to his example, the difference would have narrowed to 70 versus 50. He now estimates that open weight models reach about 90% of the efficiency of the most advanced models, and the cost per token is between 10% and 25% of the latter cost. These are Krishna’s assessments, who have not identified concrete evidence to the contrary.
It is convenient to stop at an expression that appears throughout this comparison: the open weight model. The so-called “pesos” are Numerical parameters that an artificial intelligence system learns and adapts during its training and that it uses to process information and generate feedback. When one company publishes them, others can download them and run the model on their own computers, always within the terms of their license. This does not mean we know how you trained or have access to all your data and equipment. Therefore, as the Open Source Initiative explains, having weights is not necessarily equivalent to having an open source model.
The savings that Krishna talks about are best understood by looking at how these systems are charged for use. A token is the unit with which the model represents the pieces of text it receives or generates, and many providers bill based on the amount processed. “Would you prefer something that costs 10% if it achieves 90% of the best performance, or would you rather pay ten times as much?” said the executive officer. According to IBM’s calculations, an average large company can start with two or three racks, cabinets containing servers and other equipment to run these options. But this infrastructure requires investment, electricity, maintenance and specialized personnel. Actual savings depend on usage volume, where employing an API may still be cheaper for smaller workloads.
After the cost comes an inevitable question: which tasks really need the most advanced model? Krishna acknowledges that some do, but believes that many business tasks can be solved with cheaper alternatives. He cites everyday examples, from summarizing one email and drafting another to checking a contract or taking a call from a client. “All of this can be done just as well with much cheaper options“, he assures. This is his assessment of these uses, not a guarantee that a system gives equivalent results in all cases. The choice also depends on the complexity of each task and the reliability we need.
IBM’s claims are not isolated. In May 2026, the Organization for Economic Co-operation and Development (OECD) published a report based on previous data, which placed open weighted models at about 90% of the quality index of closed alternatives when they reach the market, but cloud services cost the equivalent of 20%. There are important nuances: Stanford’s AI Index 2026 indicates that, in March, the best closed model beat Arena’s best open weighted score by 3.3%, a ranking based on user benchmarking. These are different metrics and none establish universal equality of power.
It is important to keep in mind who is making these comparisons. IBM builds its own family of Granite models and works on technologies to facilitate the deployment of artificial intelligence, although it maintains an alliance with OpenAI. So it has commercial interests in this market, but the controversy it raises goes beyond its own product. If a company can solve with a good part of its needs cheaper optionAlways using the most advanced system is no longer such an obvious decision. The question is finding the right balance between power, cost and control without compromising the quality of the results.
Picture | IBM
In Xataka | Germany troubles Google’s AI mode: When there are no links in the summary, the opinion is Google’s
