The world of artificial intelligence is undergoing a fundamental transformation. While the first wave was about who could create the biggest and most powerful model, the second wave is now focused on efficiency and sustainability. Microsoft, one of the biggest players in the industry, is currently demonstrating this trend. According to information from WION and other tech media outlets, the giant is starting to replace ChatGPT (OpenAI) and Claude (Anthropic) models with its own solutions.

The economics of tokens: Why is it hard for Microsoft to pay OpenAI?

To understand this move, it's necessary to grasp how AI services are billed. Most large language models (LLMs) operate on the basis of tokens – the basic units of text (parts of words or characters) that the model processes. For every generated token, Microsoft pays OpenAI or Anthropic. With millions of users using Copilot daily within Office 365, these costs accumulate to astronomical amounts.

Mustafa Suleyman, head of AI at Microsoft, has previously hinted that payments for external models are directly eating into the company's margins. If Microsoft wants to offer AI features as a standard part of its services, it must have full control over costs. Its own MAI (Microsoft AI) models allow the company to optimize computing power exactly where it's needed, without having to pay a "tax" to its partners.

Model comparison: Power vs. Efficiency

New models, such as MAI-Thinkin-1, are not necessarily an effort to create "the biggest model in the world." These are so-called reasoning models designed to solve complex logical tasks at much lower cost per query.

If we compare these models to the competition, we can expect the following dynamics:

• GPT-4o (OpenAI): Extremely capable, but very expensive to operate for massive integrations.
• Claude 3.5 Sonnet (Anthropic): Top-tier in programming and text nuance, but again with high per-token costs.
• MAI models: Target specific tasks (e.g., data analysis in Excel or email summaries in Outlook) with high efficiency. Their goal is to achieve similar quality to GPT-4, but at a fraction of the price.

Practical impact: What does this mean for the Czech user?

For an ordinary user in the Czech Republic or Slovakia, this change may manifest in two ways. The first is stability and speed. If Microsoft controls the entire process from the model to the interface, it can optimize application response directly for your needs.

The second key aspect is the price of services. Currently, a Microsoft Copilot Pro subscription costs approximately $20 USD per month (roughly 470 CZK), and for the enterprise segment the price is around $30 USD per user. If Microsoft succeeds in significantly reducing operating costs thanks to its own MAI models, this could in the future lead to more affordable AI features even for smaller Czech businesses or individuals.

Availability in Czech: Microsoft has a long tradition of localizing its products. It is highly likely that the MAI models will fully support the Czech language, just as the current version of Copilot does. For Czech companies, it is important to monitor whether these new models will comply with the EU AI Act requirements, which is much easier for Microsoft to achieve thanks to its own control over the technology compared to third-party models.

Trend: From scale to intelligence

This shift is not unique to Microsoft. The entire industry is moving away from endlessly increasing model parameters toward their specialization. We see this even with the Chinese model DeepSeek, which managed to offer top-tier capabilities at incomparably lower costs. Microsoft is thus trying to keep pace with a market where the winner is no longer the one with the biggest supercomputer, but the one who can generate the smartest answer with the lowest energy and cost.