GPT-5.6 is now the preferred model in Microsoft 365 Copilot

On July 9, 2026, OpenAI announced that GPT-5.6 is now the preferred model powering Microsoft 365 Copilot across Word, Excel, PowerPoint, Chat, and Cowork. The rollout covers Microsoft 365 customers using Copilot inside those apps, with Microsoft accessing the model both natively and directly through the OpenAI API.

GPT-5.6 is described by OpenAI as its latest flagship model series, positioned around getting “more useful work from every token” and offering on-demand capability for complex tasks. The company shipped the model the same day, alongside a broader ChatGPT product update and a July 8 launch called GPT-Live.

×

What shipped

The announcement includes on-record quotes from Nitin Agrawal, President of Copilot & Agents Core at Microsoft, and Nikunj Handa, Head of API Product at OpenAI. Both frame the update around producing more polished outputs in Copilot’s app surfaces: drafting in Word, analysis in Excel, presentation building in PowerPoint, and cross-functional work in Cowork.

No benchmarks, context-window numbers, tier gating specifics, or rollout regions are named in the source.

Per-app claims from the source

Surface	Stated impact
Word	Draft and edit with fewer prompting rounds
Excel	Deeper analysis, more efficient token use
PowerPoint	Rougher ideas to polished decks, less manual guidance
Cowork	Complex cross-functional work, less coordination
Chat	Inherits the same model

Why it matters

The interesting story here is not that GPT-5.6 exists. It’s that Microsoft 365 Copilot’s default brain is once again an OpenAI model, at a moment when the market had spent eighteen months speculating about Microsoft diversifying away from that dependency. “Preferred model” is careful language, and it lands hard: for the largest AI-inside-productivity deployment in the world, OpenAI is still the first pick.

This reframes two ongoing debates. First, the productivity-AI category is bifurcating. Workspace-native players like Notion AI, Canva AI, and Perplexity are each betting that AI fused into a specific work context wins. Copilot with GPT-5.6 is the same bet at Office scale, with a frontier model as the engine.

Second, the “more work per token” framing is the tell. Every serious model vendor is now pitching efficiency and quality-per-call rather than raw capability jumps, because enterprise buyers have moved past the “wow” demo and started asking about unit economics on hundreds of thousands of seats.

Microsoft is routing to OpenAI both natively and through the API, an explicit signal that the wiring between the two companies is still load-bearing regardless of Microsoft's own model work.

How it compares

The productivity-AI category splits into distinct plays, and this announcement sharpens the contrast.

• Notion AI (Notion 3.6, July 1, 2026) is embedded inside the workspace with no standalone AI product. Its bet is that workspace context is the moat and model choice is secondary. Copilot with GPT-5.6 is the same bet at a far larger installed base, but with an explicit named frontier model as the sell.
• Perplexity (Deep Research, June 19, 2026) comes at productivity from the research side and is pushing into Microsoft 365 integrations. It competes for the same seats by owning the research-and-answer layer rather than the document surface.
• Canva AI (Canva Grow 2.0, June 24, 2026) plays in creative productivity and overlaps with PowerPoint’s territory from a design-first workflow. GPT-5.6 in PowerPoint is Microsoft’s response to the “polished deck with less manual guidance” pitch Canva has been running.
• Manus AI and OpenClaw target builder and multi-channel orchestration rather than in-app productivity. OpenClaw’s multi-provider LLM routing embodies the counter-thesis: no single model should be “preferred” and workflows should route across providers.

Where this lands: Microsoft is doubling down on a single-frontier-model default at the exact moment multi-provider routing is becoming a category. That is a strategic bet, not a neutral update.

Who should care

• Copilot admins at Microsoft 365 enterprise accounts: your default Copilot model is changing. Line up an internal eval against your current prompt patterns in Word and Excel before the shift is fully live, so you can attribute any behavior changes correctly when tickets arrive.
• Platform engineering leads running parallel Copilot and third-party AI deployments: re-run your side-by-side quality evals. If GPT-5.6 closes gaps that justified a second tool, your consolidation story just got easier. If not, you have a fresh data point for the next budget cycle.
• Founders building productivity AI tools on top of Microsoft 365: assume Copilot’s baseline quality in Word, Excel, and PowerPoint just moved up. Wrappers whose value was “better drafting than Copilot” need a new wedge. Workflow-specific value (vertical templates, compliance, cross-tool orchestration) is less exposed.
• Investors and analysts watching the OpenAI-Microsoft relationship: this is a public reaffirmation of the default-model position at Microsoft’s largest AI product surface. Weight it against the diversification narrative.
• Solo operators and small teams already inside Copilot: nothing to configure. Notice whether your Excel and PowerPoint prompts start needing fewer follow-ups, and adjust your templates down.

What changed practically

Inside each Microsoft 365 surface, the model behind Copilot’s suggestions shifts to GPT-5.6. Delivery is dual-path: Microsoft serves the models natively inside Copilot and also calls OpenAI’s API directly.

The source does not specify which workloads use which path, whether the switchover is automatic for existing customers, whether admins can pin to a prior model, or which Microsoft 365 SKUs are included. No context window, latency, or benchmark numbers are provided. This is a model swap communication, not a Copilot product release. No changes to Pages, agents, or extensibility are announced.

What to watch next

1. Whether the “more work per token” claim survives real telemetry. If GPT-5.6 measurably reduces prompting rounds in Word and token spend in Excel, expect OpenAI to lean on Microsoft-sourced data in future launches. If not, expect quieter positioning.
2. How fast Cowork becomes the anchor surface. Cowork is the newest of the five apps named and the most exposed to the multi-agent orchestration pattern. GPT-5.6 landing there first suggests Microsoft is treating it as its answer to agent-native collaboration tools.
3. The diversification narrative. Making an OpenAI model the preferred default on the same day OpenAI ships it is a strong coordination signal. The next test: whether Microsoft’s own models or Anthropic’s get equivalent “preferred” placement in any Copilot surface.

Open questions

The announcement leaves the operational specifics almost entirely unaddressed:

• What is the rollout schedule? “Now the preferred model” doesn’t tell an admin whether their tenant flips this week, this quarter, or on a staged basis.
• Can admins pin to the prior model during evaluation? No fallback path is described.
• Which Microsoft 365 SKUs are included? The source says “Microsoft 365 customers” without qualification.
• What does “more useful work from every token” mean in measurable terms? No benchmarks, no before-and-after numbers, no context-window disclosure.
• How does Cowork actually behave under GPT-5.6? The description is broad enough to cover anything from smarter meeting summaries to genuine multi-agent handoff.
• What is the split between native serving and direct API calls, and does it change the data-residency and compliance picture for regulated customers? Not addressed here.

Source: openai.com