Microsoft integrates OpenAI GPT-6.2 Sol into Copilot for global users
Microsoft is rolling out new OpenAI models to Office apps while simultaneously launching its own in-house AI models to manage costs.
The short version
- GPT-6.2 Sol is rolling out to Copilot Chat, Cowork, Studio, and Office apps for eligible users.
- The heavier GPT-6.2 Astra is restricted to specific enterprise programs and requires admin activation.
- Microsoft launched its own MAI Code 2 models three days prior to target the coding market with lower costs.
- Microsoft is replacing OpenAI's image generation models in PowerPoint and Bing with its own MAI-Image-2.
- Sol costs $2 per million input tokens, while Astra costs $8 for the same input on the API.
Why it matters
This rollout shows Microsoft balancing its role as an OpenAI distributor with its goal of becoming an independent AI provider. For users, it means more powerful AI tools in familiar apps, but for businesses, it signals a shift toward managing costs via in-house models.
The new OpenAI integration
Microsoft announced on Friday, November 20, that OpenAI's GPT-6.2 Sol model is being deployed worldwide. This model will be available to eligible users across a wide range of tools, including Copilot Chat, Copilot Cowork, and Copilot Studio. It is also being integrated into the standard Office suite, specifically Word, Excel, and PowerPoint.
While Sol is the primary model for most users, Microsoft is also introducing a heavier model called GPT-6.2 Astra. Astra is designed for more complex, multi-step tasks and is currently restricted to specific environments. It reaches Copilot Cowork and Studio for all eligible users but requires a special Frontier program membership for use in Chat and Office apps.
The distinction between the two models is largely defined by their performance and price points. As Startup Fortune reported, Sol is positioned as the everyday model for common tasks like summarizing threads or drafting memos. Astra is intended for heavy-duty jobs where users might pay more for a model capable of handling longer, more intricate instructions.
How the models function
Both the Sol and Astra models operate through a system called Work IQ. This technology is designed to ground the AI in a user's specific business data, such as files, meetings, and chats. By using this framework, the AI can provide contextually relevant answers based on the information a company already possesses.
A key feature of this system is its adherence to existing security permissions. Within a company's Office tenant, the AI will provide different answers to different people based on their access rights. For example, a finance lead will see different data than an intern because the model is restricted to what each individual is already authorized to open.
From a technical standpoint, the models are being deployed with usage-based billing. This means that while some users may have access through a standard Microsoft 365 Copilot license, the more powerful Astra model requires specific administrative actions. IT teams will need to manage these settings and prepare for potential usage-limit notifications as the rollout continues.
The shift to in-house models
The timing of this rollout is notable because Microsoft recently launched its own proprietary AI models. Just three days before the OpenAI update, Microsoft introduced MAI Code 2 and MAI Code 2 Flash for GitHub Copilot. These in-house models are specifically designed to compete with OpenAI and Anthropic in the AI coding space.
Microsoft claims that its MAI Code 2 model matches 98% of the performance of GPT-6.1 Sol while costing only 40% of the price. The company also asserts that it delivers 96% of the performance of Claude Opus 5.5 at only 5% of the cost. These benchmarks suggest a strategy of using cheaper, in-house models for tasks where high-end performance isn't strictly necessary.
This dual approach allows Microsoft to keep more revenue by using its own models for certain tasks. Every request served by an MAI model generates revenue that Microsoft keeps entirely. By offering both OpenAI and MAI models, Microsoft can provide the best-known names for general office work while using its own models to cut costs in specific areas like coding.
Changing the AI landscape
Microsoft is also actively swapping out OpenAI components for its own technology in other areas. The company confirmed it is replacing OpenAI's image generation models in PowerPoint and Bing with its own MAI-Image-2. This new model currently ranks in the top three on the Arena.ai leaderboard, which tracks model performance.
This transition is part of a broader effort to reduce third-party inference costs. By moving toward its own MAI model lineup in Excel and Outlook, Microsoft aims to maintain more control over its AI ecosystem. This move follows the formal end of the exclusive partnership and revenue-sharing arrangement between Microsoft and OpenAI in April 2026.
CEO Satya Nadella has previously warned businesses against relying too heavily on foundation models from external labs like OpenAI or Anthropic. He argues that companies should build their own proprietary AI learning loops to protect their core value. This strategy creates a tension between Microsoft's massive investment in OpenAI and its push to develop independent, cost-effective AI solutions.
What to watch
- How many Copilot requests Microsoft routes to its own MAI models versus OpenAI models.
- Whether the MAI-Image-2 model becomes the standard for all Microsoft image generation tasks.
Try it
Check your Microsoft 365 Copilot settings to see if your organization has enabled the OpenAI subprocessor setting for Astra.
Sources and links
Read the original at Startup Fortune
First reported by Startup Fortune. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.