Microsoft has officially stated that GPT 5.6 is the ‘preferred model’ for Copilot, the productivity suite embedded in Office 365. The news comes amid persistent rumors of a potential breakup between the two companies, but the facts point to a collaboration that seems more solid than ever from a technological standpoint. Behind the announcement lies much more than a simple technical choice: it confirms a strategic bond that affects not only the two giants but the entire enterprise ecosystem.

Choosing a proprietary LLM as the ‘preferred’ one means, for Copilot customers, accepting a whole stack of dependencies. Every document processed, every answer generated, every automated workflow passes through a model whose evolution, controls, and costs are negotiated behind closed doors between Microsoft and OpenAI. It’s not just a performance issue: it’s an architectural decision that locks in the cloud as the sole delivery method and makes it hard for companies to evaluate self-hosted alternatives without disrupting deeply ingrained processes. For organizations with strict data sovereignty or GDPR compliance requirements, this scenario shrinks maneuvering room and shifts the control center toward the vendor.

The story takes on even sharper contours when viewed through the lens of vendor lock-in. Copilot isn’t a peripheral service: it’s deeply woven into the daily interfaces of millions of knowledge workers. Replacing the underlying model with a different one – say, an open-weight LLM running on owned infrastructure – would require not only API re-engineering but also recalibrating the assistant’s entire behavior, because characteristics of GPT 5.6 (context windows, response style, token handling) are now embedded in the application logic. This technical inertia is precisely what makes lock-in so effective.

Yet, this very move could accelerate interest in on-premise deployments. As a behemoth raises its platform barriers, more enterprises are starting to assess models like Llama 3 or Mistral, executable on dedicated hardware with sufficient VRAM, leveraging local serving frameworks and quantization techniques to optimize resources. The TCO of a self-hosted solution, compared with the per-token cost of a high-volume cloud service, can become competitive quickly, especially where cost predictability and data residency are top priorities. The AI-RADAR community, which has long analyzed these trade-offs, knows that the tipping point isn’t raw model power but the ability to keep strategic control over the infrastructure.

Breakup rumors between OpenAI and Microsoft, meanwhile, have never fully subsided. The decision to promote GPT 5.6 as the preferred model could also be read as a defensive move: a way to show the market that the axis is holding, while Microsoft continues to invest in in-house models (like the Phi family) and seeks to diversify its innovation sources. For now, though, Copilot speaks unequivocally: the best answer, according to Redmond, still comes from OpenAI’s servers.

For those shaping AI strategies, the stakes are clear: the more major vendors embed proprietary models into productivity suites, the more critical it becomes to have a plan B that runs through controllable infrastructure. It’s no longer just a simple ‘make or buy’ question; it’s about how deeply one is willing to depend on a technological supply chain that someone else defines for us.