By ByteBot
Bloomberg reported this week that Microsoft has already begun routing AI prompts in Excel, Outlook, and GitHub Copilot away from OpenAI and Anthropic models and toward its own in-house MAI models — and Microsoft AI CEO Mustafa Suleyman put the strategy in plain terms: “We pay a lot of money to Anthropic — so our goal is to reduce and ultimately eliminate that cost.” This is not a roadmap announcement. Tens of thousands of weekly prompts in Office apps are already being completed by MAI, not by GPT or Claude.
It’s Already Happening
The Bloomberg report, confirmed by TechCrunch and The Decoder on July 7, 2026, makes clear this is a live migration, not a planned one. MAI-Transcribe-1.5 is already testing in Teams. MAI-Image-2.5 is rolling out in Bing and PowerPoint. Meanwhile, Excel and Outlook are actively routing selected tasks to Microsoft’s own models.
This matters for developers beyond those using Microsoft products directly. It signals that Microsoft views OpenAI and Anthropic as cost inputs to be optimized, not strategic partners to be deepened. The partnership renegotiation in 2025 ended OpenAI’s exclusivity — this is Microsoft cashing in on that clause.
For enterprises running Office add-ins or productivity integrations that rely on specific model behaviors, the underlying model may already have changed. There was no announcement. There is no opt-out.
MAI-Code-1-Flash Is Already in Your Copilot
The most immediate developer-facing change is MAI-Code-1-Flash, a 5-billion-active-parameter model purpose-built for GitHub Copilot and VS Code. It became generally available for Copilot Business and Enterprise on June 26, and rolled out to all individual plan users (Free, Pro, Pro+, Max) through June. Individual users can select it in the VS Code model picker. Enterprise admins need to enable it via Copilot policy settings before users can access it.
Performance-wise, it beats Claude Haiku 4.5 on the tasks it was designed for: Microsoft’s official benchmarks show 51.2% on SWE-Bench Pro versus Haiku 4.5’s 35.2%, and up to 60% fewer tokens on SWE-Bench Verified. Its pricing — $0.75/M input tokens, $4.50/M output — makes it roughly three times cheaper than GPT-4o output at scale. However, it does not close the gap with Claude Opus 4.6 (80.8% on SWE-Bench Verified) or Gemini 3.1 Pro (80.6%). According to the GitHub changelog, this model suits completions, small refactors, repository Q&A, and short fixes — not complex architecture work. Microsoft is not claiming it replaces premium model routing. Not yet.
Related: GitHub Copilot Vision Is GA: Stop Describing Bugs, Show Them
The Bigger Play: MAI-Thinking-1
MAI-Code-1-Flash is a tactical move. MAI-Thinking-1 is the strategic one. Microsoft’s flagship reasoning model — 35B active parameters, sparse Mixture-of-Experts architecture with roughly 1T total parameters — scored 97.0% on AIME 2025, outpacing Claude Sonnet 4.6’s 95.6%, and was preferred over Sonnet 4.6 in a 1,276-task blind evaluation by professional raters. On SWE-Bench Pro, it reportedly matches Claude Opus 4.6.
Two caveats apply. First, MAI-Thinking-1 is currently in private preview on Microsoft Foundry, not broadly available. Second, these benchmark claims are self-published by Microsoft and have not been independently reproduced. Independent reporting places MAI-Thinking-1 roughly equivalent to Deepseek V3.2 in practice — capable, but not a clear leap over the models it intends to replace.
Consequently, if MAI-Thinking-1 continues improving — and the “hill-climbing machine” framing suggests Microsoft expects it to — the case for routing premium workloads externally weakens considerably. Microsoft holds an OpenAI license through 2032. Whether it uses it past 2027 is increasingly uncertain.
Cheaper Models, Higher Bills: The Developer Tension
The timing is worth examining. GitHub Copilot switched from flat-rate subscriptions to token-based “GitHub AI Credits” billing on June 1, 2026. The backlash was immediate: some developers on Reddit reported projected Copilot Pro+ bills jumping from $29/month to roughly $750/month on agentic workflows. A GitHub community discussion thread on the billing change accumulated over 400 comments and nearly 900 downvotes.
Microsoft is simultaneously raising the cost ceiling for developers who use premium models heavily, while replacing the models those developers relied on with cheaper in-house alternatives. Suleyman’s framing is that MAI models represent genuine capability growth. The cynical read — and it has support in the data — is that Microsoft is extracting more revenue from its developer base while cutting its own model costs. Furthermore, alternatives including Cursor, Claude Code, and OpenCode are seeing renewed interest as developers reassess their tooling.
The question isn’t whether MAI models will replace OpenAI and Anthropic inside Microsoft products. That’s already happening. The question is how fast, and whether MAI quality climbs fast enough that developers stop noticing.
Key Takeaways
- Microsoft is actively routing Office app prompts to MAI models as of July 7, 2026 — this is live, not planned
- MAI-Code-1-Flash is now in the GitHub Copilot model picker for all plans; it beats Claude Haiku 4.5 but trails Opus-class models and suits completions and small refactors best
- MAI-Thinking-1’s headline benchmarks are self-published and not independently verified, placing it roughly in the Sonnet-class range
- Suleyman explicitly stated the goal is to eliminate Anthropic payments entirely — this is strategy, not implication
- The concurrent Copilot token-billing shift means developers pay more while Microsoft cuts its own costs — alternatives like Cursor and Claude Code are gaining renewed attention
