Microsoft Cancels Internal Claude Code, Launches 7 MAI Models at Build 2026

Multiple reports pointed to cost as the real driver . The fiscal-year-end timing strongly suggested financial motives , and the move eliminated what had become an expensive line item. An anonymous source told one outlet the tool had become "too expensive" . The broad Microsoft AI CEO Mustafa Suleyman quote about white-collar job automation within 18 months that is sometimes attached to this story comes from a separate Fortune interview about AI automation broadly, not a specific Claude Code pricing remark .

The cost numbers that triggered the reckoning

AI coding tools that bill by token consumption can get expensive fast. The standout cautionary tale is Uber, which reportedly rolled out Claude Code to its engineering teams starting in late 2025. Per-engineer monthly costs ran between $500 and $2,000, and the company exhausted its entire $3.4 billion 2026 AI budget in four months, according to Fortune reporting cited in multiple sources .

Another unnamed company reportedly racked up roughly $500 million in Claude Code charges in a single month because no one had set a usage cap for employees .

For Microsoft's internal use, the math was simpler and equally concerning: an AI coding tool that costs $500 to $2,000 per engineer per month becomes a significant budget item when thousands of engineers are using it. At the high end, $2,000/month in Claude Code API costs added roughly $24,000 per year per engineer—an 8-12% surcharge on existing headcount costs with no corresponding headcount reduction .

Microsoft's response: seven in-house MAI models

At Build 2026, held June 2–3 at the Fort Mason Center in San Francisco, Microsoft AI CEO Mustafa Suleyman took the stage to unveil seven new in-house AI models under the unified MAI (Microsoft AI) brand . All seven were developed by the Microsoft AI Superintelligence Team and trained from scratch on commercially licensed data with zero distillation from third-party models, including OpenAI's GPT series .

The complete MAI model family covers reasoning, coding, image generation, speech recognition, and speech synthesis :

• MAI-Thinking-1 – Microsoft's first in-house reasoning model, a 35-billion-active-parameter Mixture-of-Experts system with a 256,000-token context window
• MAI-Code-1-Flash – A 5-billion-parameter coding model that started rolling out to all paying GitHub Copilot users on day one of Build
• MAI-Image-2.5 and MAI-Image-2.5-Flash – Image generation models
• MAI-Transcribe-1.5 – Speech recognition and transcription
• MAI-Voice-2 and MAI-Voice-2-Flash – Speech synthesis

Benchmark positioning against Claude

Microsoft positioned MAI-Thinking-1 aggressively against Anthropic's models. The key claims, as presented at Build 2026:

• 97.0% on AIME 2025 and 94.5% on AIME 2026, two of the hardest mathematical reasoning benchmarks in public circulation
• Matches Claude Opus 4.6 on SWE-Bench Pro (53% score), a software engineering benchmark that measures how well an agent resolves complex, multi-step coding tasks
• Independent blind raters preferred MAI-Thinking-1 over Claude Sonnet 4.6 in human evaluation tests

MAI-Thinking-1 doesn't top the overall leaderboard, and Microsoft isn't pretending it does . But in the medium-size weight class, the company says it's punching above its weight, with a smaller inference footprint than dense models of comparable capability . Microsoft emphasized low token cost as a key advantage for enterprise use .

MAI-Code-1-Flash scored 51.2% on SWE-Bench Pro in testing versus Claude Haiku 4.5's 35.2% in the same VS Code harness, according to Microsoft engineering posts . It began rolling out across Visual Studio Code and GitHub Copilot immediately .

The apparent contradiction: selling Claude while cutting Claude

Here's the tension that makes this story interesting: at the same time Microsoft was canceling internal Claude Code licenses, Anthropic was gaining market momentum. In U.S. business AI payments tracked by Ramp, Anthropic reportedly passed OpenAI for the first time, growing from 9% of businesses paying for Anthropic products in May 2025 to 35% a year later .

And Microsoft itself is far from abandoning Anthropic commercially. In March 2026, Microsoft announced Copilot Cowork—built in close collaboration with Anthropic, using Claude's agentic model for multi-step tasks inside Microsoft 365 apps . Copilot Cowork launched inside the new Microsoft 365 E7 "Frontier Suite" tier at $99 per user per month, a 65% jump from the $60 E5 tier . That bundle also includes Copilot, Agent 365, and the Entra identity suite .

The through-line is consistent when you separate internal tooling from commercial partnerships. Internally, Microsoft is reducing direct reliance on Claude Code where it was cutting into margins without corresponding headcount savings . Commercially, it's packaging Claude's capabilities into premium enterprise subscriptions at profitable markups. Claude models also remain available through Copilot CLI and in Microsoft's Foundry platform, so the internal licensing cancellation is not a full break with Anthropic .

The broader enterprise AI cost reckoning

What happened inside Microsoft is part of a larger shift in how enterprises think about AI tooling access. The era of "just give everyone an AI coding tool and see what happens" is giving way to a more managed phase where access, tooling, and model choice are tightly controlled .

This isn't just about Claude Code. Anthropic itself acknowledged the cost reality by trimming its 2025 gross profit margin forecast to around 40%, roughly ten percentage points lower than earlier expectations, citing higher-than-planned cloud compute costs . In April 2026, Anthropic restructured enterprise Claude pricing toward usage-based billing with higher enterprise minimums and per-seat charges tied to consumption commitments .

Enterprise buyers are now operating in a world where per-engineer AI coding tool costs routinely reach $500 to $2,000 per month , and one rogue deployment without proper usage caps can burn through half a billion dollars . The response from large organizations is increasingly clear: consolidate onto platforms they control, build in-house alternatives, and only pay the premium for external models when those models are on the revenue-generating side of the business. Microsoft just demonstrated all three.

Suleyman's broader strategic framework—which he calls "self-sufficiency"—is about building models without depending on any single external provider . That doesn't mean breaking with OpenAI or Anthropic; the partnerships remain. But the Build 2026 MAI launch and the internal Claude Code pullback together signal that Microsoft intends to be its own primary model provider by 2027 .