OpenAI wants you to believe the marriage is intact. On Thursday, the company crowned GPT 5.6 the “preferred model” for Microsoft 365 Copilot — Word, Excel, PowerPoint, and the new Cowork app. The phrasing is deliberate. “Preferred” sounds like a victory lap. It also sounds like a lawyer’s compromise.
Bloomberg had reported days earlier that Microsoft is quietly swapping OpenAI tech for its own MAI models inside those same apps. The motive: cost. The implication: a slow divorce. OpenAI’s blog post doesn’t deny the swap. It simply asserts a continued role. That role, however, is undefined. “Preferred” could mean default, primary, or merely first among equals. Microsoft hasn’t echoed the language. Silence is its own statement.
Large language models are expensive. Microsoft knows this better than most; it runs the infrastructure. Every token generated by OpenAI’s API is a line item on Azure’s bill. Building MAI in-house lets Microsoft amortize training spend across its own cloud, control latency, and avoid per-call fees. The economics favor insourcing. OpenAI knows this too. Its “preferred model” claim is a retention play — keep the volume high enough to justify the partnership’s pricing tier.
But the partnership was always asymmetric. Microsoft invested billions for early access and integration rights. OpenAI got compute and distribution. Now each side has leverage the other needs. Microsoft has the enterprise install base. OpenAI has the model frontier. The tension is structural, not personal.
“Our partnership with Microsoft has always been about bringing the benefits of advanced AI to more individuals and organizations, and we’re excited to continue building on that shared commitment.” That sentence is crafted for regulators, investors, and employees. It says nothing about exclusivity, revenue split, or roadmap alignment. It says “we’re excited.” Excitement is not a contract.
The blog post also lists Cowork — a brand so new it barely exists in public discourse — as a showcase. That feels like a signal: OpenAI wants credit for the next thing, not just the legacy apps. Microsoft may have other plans for Cowork’s intelligence layer.
In practice, “preferred model” likely means: when Copilot needs a general-purpose reasoning engine, it calls GPT 5.6 first. Specialized tasks — spreadsheet formula prediction, slide design heuristics — may route to MAI. Microsoft gets best-of-breed per workload. OpenAI gets volume on the high-margin, high-visibility queries. Both sides can claim victory.
But the arrangement is fragile. If MAI improves fast enough, the “preferred” label becomes a courtesy. If GPT 5.6 stumbles on hallucination or latency, Microsoft will reroute without announcement. The partnership survives only while mutual convenience exceeds mutual friction.
Reporters keep asking about a split because the signals are mixed. Microsoft’s GitHub Copilot still leans on OpenAI. Azure OpenAI Service sells OpenAI models to Microsoft’s own customers. Yet Microsoft Research publishes papers on phi-3, Orca, and MAI — models that compete directly. The company is hedging. OpenAI is hedging too, selling API access to Salesforce, Shopify, and anyone else with a checkbook.
This isn’t a breakup. It’s a renegotiation conducted in public. The “preferred model” label is a press release, not a protocol spec. Watch the API call logs, not the blog posts. They’ll tell you who actually powers the next Word suggestion.
Enterprise buyers should ignore the branding. They need to know: which model handles their proprietary data, what the latency SLA is, and who owns the fine-tuning pipeline. “Preferred” answers none of that. Microsoft’s product managers will decide per feature. OpenAI’s sales team will fight for each slot.
The industry should stop treating this as a romance. It’s a supply chain. Supply chains shift when economics change. The announcement this week changed nothing fundamental. It just bought both companies a few more months of comfortable ambiguity.