The head of Microsoft AI aims to cut expenses related to Anthropic.

      Mustafa Suleyman has pinpointed Microsoft’s main competitor in the AI sphere, and it isn’t OpenAI. In an interview with Bloomberg, the head of Microsoft’s internal model initiative stated, “Anthropic is very costly, and I believe many are actively seeking alternatives.” This assertion goes beyond mere competitive strategy; it signifies a clear objective. “We spend a significant amount on Anthropic, so our aim is to minimize and ultimately eliminate that expense,” Suleyman noted.

      To support this claim, Microsoft unveiled seven new in-house AI models at its annual Build conference for developers, including MAI-Thinking-1, a reasoning model that it claims performs on par with Anthropic’s Claude Opus 4.6 on a widely accepted coding benchmark but at a lower cost.

      The pricing argument comes at a time when expenditures on enterprise AI are proving to be a pressing issue. Uber exhausted its entire AI coding budget for 2026 within just four months and has since implemented a monthly limit of $1,500 per employee for each tool. Similarly, Walmart has restricted access to its internal AI assistant after usage surpassed expectations.

      For Microsoft, which heavily relies on AI tokens for Copilot and its engineering teams, the cost concern is critical. Suleyman mentioned, “Numerous individuals within our organization are spending millions of dollars” on AI tokens. Developing affordable in-house alternatives isn’t only a strategy against external competitors—it’s also a means of safeguarding Microsoft’s own profit margins.

      After refining its models for consulting firm McKinsey, Microsoft claims it has achieved ten times the cost efficiency compared to OpenAI’s GPT 5-5, as noted by Suleyman. According to Bloomberg, the company has also been in discussions with Adobe about utilizing its internal models.

      Until late 2025, Microsoft was contractually restricted from independently pursuing advanced AI development due to its partnership with OpenAI. A renegotiated agreement has since allowed the company to create competing models while retaining licensing rights to everything OpenAI develops until 2032.

      That timer is now running. Suleyman’s MAI Superintelligence team, established in November 2025, has delivered its first public models within six months. Although this pace is aggressive for a company relatively new to cutting-edge model development, a gap still exists.

      Suleyman was frank regarding this gap. "We’ve closed a significant gap in six months," he told Benzinga, while recognizing that Anthropic has introduced two more advanced models since Opus 4.6, affording it several months of advantage.

      The strategic analysis is illuminating. Microsoft has discounted access to OpenAI’s models until 2032 but lacks a similar agreement with Anthropic. This makes Anthropic a more costly dependency and a target for Suleyman to address with in-house options.

      There’s also a competitive rationale. Anthropic’s Claude models have become the preferred choice for enterprise AI coding tools, and the company is gearing up for an IPO that could value it at over $1 trillion. If Microsoft can provide comparable performance at a lower cost, coupled with the Azure infrastructure already in use by its enterprise clients, it diminishes the need for purchasing Anthropic separately.

      Whether Microsoft’s models can truly rival Anthropic’s latest versions, rather than an earlier one, remains an unanswered question with Suleyman’s new models. While matching Opus 4.6 on a benchmark is a solid starting point, Anthropic has already surpassed that milestone. In a rapidly evolving market, where advancements occur every few months, keeping pace and maintaining that position are distinctly different challenges.