Suleyman Says Anthropic is "Extremely Expensive" as Microsoft Accelerates Plans to Eliminate Dependency

Mustafa Suleyman, CEO of Microsoft AI, provided last Thursday to Bloomberg the most straightforward line an executive from Redmond has produced about its largest third-party model supplier. "Anthropic is extremely expensive, and many people are urgently looking for alternatives," Suleyman said. "We pay a lot of money to Anthropic, so our goal is to reduce and ultimately eliminate that cost."

The statement came two days after Microsoft Build 2026, where Satya Nadella unveiled seven proprietary models built without distillation from OpenAI or Anthropic. Among them is MAI-Thinking-1, a reasoning model with 35 billion active parameters in a sparse Mixture of Experts architecture with around 1 trillion total parameters and a context window of 256,000 tokens. Microsoft claims 97.0% on the AIME 2025 and 94.5% on the AIME 2026, benchmarks for mathematical reasoning where today's competitive frontier models score in the high range.

The urgency is financial. Microsoft is committing between $120 billion and $130 billion in AI capex for fiscal year 2026. A significant portion of this flow powers data centers that run Claude for Foundry clients. Reducing third-party model licensing bills is the quickest way to open margins within a cost structure that has already reached Wall Street's comfort limit.

Pressure Hits Corporate Buyers

For the CIO of a European insurance company or an American health provider, the operational message is clear: the price of frontier tokens will drop, and vendor lock-in with a single model provider has become the most expensive decision one can make in 2026.

Capgemini, Cognizant, and Wipro standardized part of their delivery centers on Claude Opus 4.7 in the last twelve months. They now have to defend that choice in procurement committees. Microsoft has listed over 11,000 models on Foundry, including Claude Opus 4.8, and introduced an automatic Model Router: the infrastructure selects the cheapest model that meets the customer's SLA. This routing alone reduces costs in code generation use cases by 30% to 50%, according to benchmarks Microsoft released during a conference with analysts on June 2.

Anthropic did not publicly comment on Suleyman's statement by the time this article was published. Dario Amodei, in an interview with the New York Times last month, argued that Anthropic will price according to the value it delivers, not according to the pressure from competitors that have not yet reached the same capability level. The company also has the hyperscaler side: Amazon has invested $25 billion cumulatively, and Google is the second-largest inference customer.

The Outlook Outside the United States

The price war is not just an American issue. In India, TCS, Infosys, and Wipro have just surpassed 300,000 combined Microsoft 365 Copilot licenses in deployment, according to data released by Microsoft on June 3. Each cost reduction in the underlying Copilot model translates into direct margin for these firms, which price transformation projects with embedded Copilot on a fixed-fee basis. The ability to prove that the backend Copilot would run on a proprietary Microsoft model, and not on Claude, opens up negotiation conversations with the end customer.

In the United Kingdom, where Lloyds Banking Group and HSBC piloted Foundry in May to integrate agents into KYC workflows, the Model Router changes the TCO equation by removing the "which model" decision from the architect's hands and placing it into an internal bidding system among providers. European banks operate within DORA, and auditing a single model provider has become a governance item that multiplexing resolves.

In Japan, MUFG and Mizuho are in the final stages of selecting a large language model provider for their call center. Microsoft is currently competing against NEC and NTT Data for this contract, and the argument of "eliminate the payment to Anthropic via Model Router" weighs differently for Japanese boards that historically prefer a single vendor ecosystem.

The Unspoken Piece

Suleyman was a co-founder of DeepMind and Inflection AI before Microsoft acquired his team in March 2024. The phrase "eliminate the cost" carries personal weight. Microsoft paid $650 million to bring him and the Inflection team aboard, and the expected return on investment is measured in how many percentage points of Anthropic's bill he can replace with Microsoft code. The cycle is far from closing. Anthropic remains one of the largest AI suppliers to Foundry, and existing contracts extend until 2028.