Microsoft Aims to Reduce Token Costs by Developing In-House AI Models

According to Beating Monitor, Mustafa Suleyman, Microsoft’s AI lead, stated bluntly in an interview with Bloomberg that Anthropic’s models are extremely expensive, prompting many users to urgently seek alternatives. Suleyman revealed that Microsoft currently pays substantial sums to Anthropic, and the team’s clear objective is to reduce and ultimately eliminate these procurement costs. To reduce dependence on external frontier models, Microsoft unveiled seven self-developed MAI-series models at the Build 2026 developer conference in June. The newly released models include MAI-Thinking-1, a model with reasoning capabilities. Microsoft claims these new models significantly lower token costs while maintaining performance and can compete with Anthropic’s flagship model, Claude 4.6 Opus, on tasks such as programming. The urgency to control costs has prompted Microsoft to restructure its internal toolchain: the company is gradually reducing reliance on third-party tools like Anthropic’s Claude Code for programming and migrating internal development and daily workflows to its self-developed GitHub Copilot CLI to manage internal token consumption. Microsoft’s accelerated development of proprietary models follows the restructuring of its partnership with OpenAI. On April 27, 2026, both parties signed a revised agreement that ended Microsoft’s exclusive license to OpenAI’s technology and terminated Microsoft’s revenue-sharing payments to OpenAI. This restructuring clears the path for Microsoft to independently develop frontier models and directly compete with OpenAI, Google DeepMind, and Anthropic.