Microsoft announces model controls and a voice model at Ignite


Microsoft says MAI Licensed Weights will let organizations customize and deploy its models in Azure. It also introduced MAI Realtime 1 for enterprise voice conversations, with pricing advertised from $0.01 per audio minute.
Microsoft announced MAI Licensed Weights and MAI Realtime 1 alongside its MAI Code 2 coding models, extending its Ignite offerings beyond programming. The company describes the licensed models as offering frontier performance and says the voice model is designed for enterprise conversations, with pricing starting at $0.01 per audio minute.
The announcements build on Microsoft’s earlier work adapting models for individual customers. MAI Code 2 and Code 2 Flash are generally available on GitHub; Flash is also in private preview in Microsoft Foundry.
Earlier: Microsoft unveils MAI Code 2 coding model at Ignite
Microsoft says Licensed Weights will let organizations customize models with their own knowledge and run them in their own Azure environments. Customers will control deployment, access and upgrade timing. Their data and adaptations will remain theirs, and Microsoft will use them to train its models only with their express agreement. The company says those controls are intended to help businesses maintain continuity as they adapt models.
Microsoft says Licensed Weights is coming soon through its Frontier Company engagements and directs interested organizations to their Microsoft account teams. It has not disclosed a launch date or detailed licensing and commercial terms.

Microsoft says the programming model nearly matches rivals at a fraction of the cost. The company also released image-generation and transcription models and plans limited access to a cybersecurity model.
Microsoft says its customer-specific evaluations showed the following results for adapted models. The company cautions that performance and cost advantages vary by workload, model configuration and evaluation method.
| Customer and task | Result reported by Microsoft |
|---|---|
| AT&T: 5G network analysis | Customized models outperformed GPT-6.1 Sol on standard tasks at 1/50th of the cost; Microsoft describes performance on more complex tasks as frontier-level at significantly lower cost. |
| Woven by Toyota: code fixes | An adapted Flash model produced ready-to-merge fixes in 44% of cases where existing models needed human intervention, at less than one-tenth of the cost. |
| Alaska Airlines: maintenance-report classification | Accuracy rose from 81.2% to 91.1% after tuning; qualified employees retain responsibility for final decisions. |
MAI Realtime 1 is Microsoft’s first speech-to-speech model: it is intended for applications that conduct spoken conversations, with pricing starting at $0.01 per audio minute. Microsoft says it outperformed GPT Realtime 2.1 on quality, speed and price across public benchmarks and its own product evaluations. It also says the model sits on the performance-price Pareto frontier for Big Bench Audio, a speech-reasoning test, and Tau-Voice, a task-completion test.
For coding, Microsoft says Code 2 averaged 98% of GPT 6.1 Sol’s performance at 40% of its cost across four everyday enterprise coding benchmarks, and cost 5% as much as Opus 5.5. It says Flash matched GPT 6 Luna and Haiku 5.5 on everyday coding tasks at 65% and 20% of their respective prices. Microsoft says the coding models are not yet tuned for fundamental science research or very long-running autonomous work.
Microsoft also made Word, Excel and PowerPoint generally available inside Copilot. Users can create documents, spreadsheets and presentations there and edit the files in their respective Office applications. Microsoft IQ, which includes a system called Work IQ designed to help agents understand an organization’s communications and work, also became generally available. CEO Satya Nadella described Microsoft IQ as an externalization of the context system underlying Copilot, intended to help organizations build, retain and continually improve their intelligence.
A new Copilot interface called Home combines chat with Cowork, where users can delegate tasks to superintelligence (SI) agents. Microsoft says Home’s Auto feature routes requests to suitable models and can recommend moving complex work to Cowork with the user’s agreement. Today, a feature in private preview, scans email, calendars and Teams threads to highlight tasks and can prepare replies or suggest schedule changes, SiliconANGLE said.
Copilot Autopilot, a cloud-based agent available to subscribers in Microsoft’s Frontier early-access program, can be authorized to work across services such as email and calendars. Microsoft says it can automate recurring processes and manage multistep work such as supplier reviews. The company says each Autopilot agent will have a unique ID, administrators can restrict or block agents, their activities are logged for human review, and agents can ask for approval before certain actions. Billing will depend on usage through Copilot Credits; final pricing has not been announced.
CEO Satya Nadella also announced Copilotforce, a partnership with Salesforce intended to bring Salesforce context into Copilot, Teams and Fabric. Microsoft says it has 38 prebuilt skills for sales tasks, including retrieving customer data and updating customer records.
Copilot Workspaces, a shared interface for teams working with a model, is available through Frontier early access; Microsoft plans to let users bring personal Autopilot agents into it. Its Decisions 1.1 model is generally available. Microsoft also plans to launch Copilot Code in a few weeks, allowing users to request dashboards, desktop widgets and small business applications in everyday language.
6 min