Microsoft Launches Copilot Cowork with Usage-Based Pricing and Unveils MAI-Thinking-1 Reasoning Model

Microsoft has made two significant announcements that are reshaping the enterprise AI landscape. First, Copilot Cowork an autonomous AI agent capable of independently executing complex, multi-step tasks across Microsoft 365 has gone generally available with a new usage-based pricing model. Second, the company unveiled MAI-Thinking-1, its first in-house reasoning model, marking a strategic pivot toward building foundational AI capabilities in-house. These developments underscore Microsoft's commitment to delivering powerful, flexible, and cost-effective AI solutions while reducing reliance on third-party models.

Copilot Cowork: Autonomous AI for the Enterprise

Initially unveiled in March 2026, Copilot Cowork represents a leap beyond simple chatbots. It is designed to act on users' behalf, automating workflows across Microsoft 365 applications like Word, Excel, PowerPoint, Teams, and Outlook. Powered by Work IQ Microsoft's intelligence layer built on the Microsoft Graph the agent can reason over emails, meetings, and files, and interact with business systems such as Fabric, Dynamics 365, and a growing ecosystem of plugins.

How It Works

Users describe a desired outcome, and Copilot Cowork plans, executes, and delivers real work. For example, a user could ask: "Draft a status report for the Q2 project, summarize the key risks from the latest emails, and schedule a review meeting with the team." The agent then autonomously performs these steps, leveraging context from the user's work data.

Copilot Cowork supports multiple AI models at launch, including Anthropic's Opus 4.8 and Sonnet 4.6. Frontier program users also have access to OpenAI's GPT 5.5. Microsoft plans to introduce its own Azure-hosted model, "Cowork 1," optimized for cost, quality, and enterprise-grade use, including bias removal. The company is also exploring integrating an open-source model from DeepSeek to further reduce costs.

Usage-Based Pricing: A Shift to Consumption

Perhaps the most notable change is the move from flat per-user fees to usage-based pricing. Billing is denominated in "Copilot Credits," with the price of each task determined by four factors: model use, context retrieval, tool calls, and runtime. More intensive tasks those requiring deep reasoning, multiple data sources, or generating several outputs cost more credits.

Customers have two payment options:

Pay As You Go: $0.01 per Copilot Credit.
P3 (Pre-paid, Provisioned, Priority): Commit to a volume of credits in advance for a discount.

Importantly, Copilot Cowork requires an existing Microsoft 365 Copilot user license ($30/user/month for enterprises, $20 for business). Microsoft claims that, on average, Copilot Cowork is 30-40% cheaper than using Claude Cowork with its Microsoft 365 connector.

Enterprise Readiness and Cost Management

IT administrators can manage costs by controlling when and to whom Cowork is available, imposing spending limits at tenant, group, and user levels, and receiving notifications. Users see the credit cost of each task before execution. All actions are auditable, and the agent runs in a protected, sandboxed cloud environment within Microsoft 365's security and governance frameworks.

MAI-Thinking-1: Microsoft's First In-House Reasoning Model

Announced at Build 2026 on June 2, 2026, MAI-Thinking-1 is the flagship of a new family of seven MAI models. It is a true reasoning model that breaks down complex problems into multiple steps before generating an answer, similar to reasoning modes from other leading AI labs.

Technical Architecture

MAI-Thinking-1 is a medium-sized, sparse Mixture of Experts (MoE) model with 35 billion active parameters and approximately 1 trillion total parameters. This sparse architecture activates only a fraction of the network for each token, resulting in a smaller inference footprint and lower operational costs compared to dense models of similar capability.

Key differentiator: The model was trained from scratch on clean, commercially licensed data without distillation from third-party models. Microsoft explicitly excluded AI-generated content from its pre-training datasets, arguing that distillation limits generalization. This approach emphasizes genuine reasoning over imitation. The training leveraged Microsoft's "hill-climbing machine," an integrated process for continuous empirical improvement.

Context Window and API Compatibility

MAI-Thinking-1 boasts a 256K token context window, capable of processing documents up to 600 pages long. It is compatible with the widely used Chat Completions API and supports function calling, making it easy to integrate into existing applications.

Benchmark Performance

MAI-Thinking-1 demonstrates strong performance across several benchmarks:

Software Engineering: Matches Claude Opus 4.6 on SWE-Bench Pro.
Mathematical Reasoning: Achieved 97.0% on AIME 2025 and 94.5% on AIME 2026.
Human Evaluation: Preferred over Sonnet 4.6 in blind side-by-side evaluations for overall quality across single and multi-turn tasks.

Microsoft describes the model as a "medium-sized model that stands among the strongest in its weight class" and a step toward "Humanist Superintelligence: advanced AI designed to serve people and organizations rather than replace them."

Strategic Implications

These announcements signal Microsoft's intent to build frontier intelligence on its own terms. By developing its own models and offering flexible pricing, Microsoft reduces reliance on OpenAI and other third-party providers, gaining more control over the enterprise AI stack. The usage-based pricing model reflects the increasing computational demands of AI workloads, offering customers a pay-for-what-you-use approach that aligns billing with actual value.

For enterprises, the combination of Copilot Cowork and MAI-Thinking-1 provides a powerful, integrated platform for automating complex tasks with state-of-the-art reasoning capabilities. As Microsoft continues to expand its AI ecosystem, these innovations are likely to drive further adoption of AI in business processes.

Getting Started

To try Copilot Cowork, you need an existing Microsoft 365 Copilot license. IT administrators can enable the agent through the Microsoft 365 admin center and configure spending limits. Developers can access MAI-Thinking-1 via the Azure AI Foundry portal and the Chat Completions API, with documentation available on Microsoft Learn.

As these technologies evolve, expect deeper integration, more cost-effective options, and continued improvements in reasoning and automation capabilities. Microsoft is clearly betting on a future where AI agents and reasoning models become central to enterprise productivity.