Microsoft presented seven new AI models from its own MAI division at its Ignite conference. They cover the areas of code, security, images, speech and routing. Routing here means distributing requests to the appropriate model in each case, instead of sending every task to the largest and most expensive system.
What is remarkable is less the number of models than the benchmark: Microsoft explicitly measures itself against the models of its own partner OpenAI. For MAI Code 2, the company states 98 percent of the performance of GPT-6.1 Sol at 40 percent of the cost. These are manufacturer figures; independent reviews are still pending.
The cost factor moves to the foreground
The message behind these figures is clear: the focus is not on the peak performance of individual models, but on the ratio of output quality to ongoing costs. This is a shift compared with previous years, in which above all benchmark records were communicated.
The reason for this lies in the way AI is now used. A single chat dialogue generates manageable costs. An AI agent, i.e. a system that works through tasks independently in several steps, generates a multiple of model calls for the same task: planning, calling tools, checking results, refining. If such an agent runs continuously in a business process, the price per call decides the economic viability of the entire undertaking.
Continuous operation instead of single queries
Microsoft names precisely this continuous operation of AI agents in companies as the focus of the new models. In doing so, the company addresses a point at which many pilot projects currently get stuck: the proof of concept works, but the extrapolation to regular operation fails because of token costs.
A difference of 60 percent in costs changes this calculation noticeably. Use cases that were previously marginally uneconomical – such as the automatic pre-screening of every incoming ticket, the continuous analysis of code changes or the ongoing monitoring of document inventories – come back within reach.
What the difference of two percentage points means
The figure of 98 percent of the performance sounds like a negligible gap. In practice, it depends heavily on the use case where this gap shows up. For routine tasks with a clear structure, it usually does not carry much weight. For complex, multi-stage tasks, a small quality difference can add up over many steps, because every error burdens the next step.
In practice, this does not mean an either-or, but a division. The combination of inexpensive working models for mass business and a stronger model for critical decisions is currently the obvious pattern. The fact that Microsoft includes its own routing model in the package fits this logic: selecting the right model itself becomes a technical task.
Strategic level: Microsoft alongside OpenAI
The fact that Microsoft pits its own models directly against those of OpenAI is interesting for users above all for one reason: it points to more choice within an ecosystem that many companies use anyway. At the same time, it shows that vendor relationships in this market remain in motion.
For one's own architecture, this means less committing to a single vendor now and more keeping interchangeability constructive. Anyone who hard-codes model calls into the application logic today pays twice with every switch – once for the migration and once for the renewed quality assurance.
Practical consequences for ongoing projects
- Measure cost per transaction, not per request. What matters is what a fully processed business transaction costs, not the list price per million tokens.
- Encapsulate model access. A separate layer between application and model makes switches and A/B comparisons possible without touching the business logic.
- Build your own test cases. Public benchmarks say little about your own domain. A collection of real cases from your own operations is more meaningful than any percentage figure from the manufacturer.
- Tier instead of standardise. Not every subtask needs the strongest model. Deliberate tiering often lowers costs more noticeably than switching vendors.
- Treat manufacturer figures as a hypothesis. The stated values come from Microsoft itself and are a reason to recalculate, not a basis for a budget.
Assessment
For web and software projects, the central question shifts from "Which model is the best?" to "Which model is cheap enough for this step?". Anyone who wants to embed AI agents permanently in processes or development pipelines should therefore plan the choice of model as an interchangeable component of the architecture and underpin it with their own test cases and a cost calculation per transaction. If the price of solid models continues to fall, it is no longer the technology that decides the benefit, but the quality of the processes into which it is embedded.