GPT-5.6 in Copilot: Cost control comes first

Microsoft and OpenAI are rolling out GPT-5.6 in Microsoft 365 Copilot. According to Microsoft, the model is becoming available in Word, Excel, PowerPoint, Copilot Chat and Copilot Cowork, and Copilot may use it automatically when it is better suited to a task.

What changes now

Microsoft describes GPT-5.6 as the preferred model for Microsoft 365 Copilot. The emphasis is not only on better answers, but on multi-step knowledge work: structuring documents, condensing Excel analysis, drafting presentations and using Cowork to execute tasks across files, tools and business context.

In parallel, OpenAI names three model tiers: Sol as the highest-capability model, Terra as the balanced option and Luna as the cost-efficient tier. For the API, OpenAI lists prices per 1 million tokens: Sol at 5 US dollars input and 30 US dollars output, Terra at 2.50/15 US dollars and Luna at 1/6 US dollars. OpenAI also says prompt caching keeps the 90 percent discount on read cached inputs; cache writes cost 1.25 times the normal input rate.

Why this matters for enterprises

For CIOs, the Copilot model change is not just a feature update. As Office work, chat and Cowork become more agentic, the control question shifts: Which tasks may automatically use a more expensive model? Where is a cheaper model sufficient? Which teams need maximum quality, and which need predictable unit economics?

That makes Copilot governance more operational. It touches policies for model choice, data access, cost allocation, logging and traceability of results. Especially in regulated DACH organizations, it must be clear which data is processed in which context and who is accountable for approved workflows.

DACH perspective

The practical starting point is not a broad rollout of every new capability. A small catalogue of prioritized use cases is more useful: management reporting, proposal documents, project status, contract analysis or Cowork processes. For each use case, quality, data-protection risk, review duty and cost logic should be defined.

The guiding question is simple: Do you treat GPT-5.6 in Copilot as a new model — or as a new budget and governance layer for knowledge work?