Microsoft’s own reasoning model, MAI-Thinking-1, has been in public preview on Microsoft Foundry since August 12, 2026. Does it reduce Microsoft’s reliance on OpenAI? Microsoft’s release, benchmark table and product posts give a narrower answer than the headlines. The contract timeline lives in our contract explainer.
The contract backdrop, in brief: since April 27, 2026 Microsoft’s license to OpenAI’s IP is non-exclusive and runs through 2032, and OpenAI’s revenue-share payments to Microsoft continue through 2030 at the same percentage, subject to a total cap (Microsoft, April 27).
This post asks something narrower: what does Microsoft’s own model release say MAI-Thinking-1 can and cannot do next to OpenAI’s models, and how large is the dependence in Microsoft’s filings? It is news analysis, not investment advice.
Key Takeaways
- Microsoft’s license to OpenAI’s IP became non-exclusive on April 27, 2026; the full timeline is in our contract explainer.
- In Microsoft’s own benchmark table, MAI-Thinking-1 trails GPT 5.4 on all three benchmarks where both report: GPQA Diamond 84.2 vs 92.8, Terminal Bench 2.0 46.0 vs 75.1, SWE-Bench Pro 52.8 vs 57.7.
- Microsoft says MAI-Thinking-1 was trained “without distillation from third-party models” and never names OpenAI. The product results it has published (Excel, GitHub Copilot) come from models derived from MAI-Code-1-Flash, not MAI-Thinking-1.
- Microsoft’s FY2026 10-K reports $24.1 billion of revenue from commercial arrangements with OpenAI, about 7.3% of its $331.8 billion total revenue (our calculation). Bloomberg estimates about 70% of Microsoft’s AI sales; that is an estimate, not a Microsoft figure.
The Contract Backdrop in One Paragraph
Microsoft’s April 27, 2026 post says Microsoft remains OpenAI’s primary cloud partner, OpenAI products ship first on Azure unless Microsoft cannot and chooses not to support them, and OpenAI can serve customers across any cloud provider.
For the October 2025, February 2026 and April 2026 changes in order, see our contract explainer.
What MAI-Thinking-1 Can and Cannot Replace, by Microsoft’s Own Numbers
Microsoft’s MAI-Thinking-1 release includes a benchmark table with eight benchmarks. GPT 5.4 is the only OpenAI model in it, and it has scores for three of the eight. The table below lists those three; the gap column is our subtraction.
| Benchmark (Microsoft’s table) | MAI-Thinking-1 vs GPT 5.4 | Gap (points) |
|---|---|---|
| GPQA Diamond | 84.2 vs 92.8 | GPT 5.4 ahead by 8.6 |
| Terminal Bench 2.0 | 46.0 vs 75.1 | GPT 5.4 ahead by 29.1 |
| SWE-Bench Pro | 52.8 vs 57.7 | GPT 5.4 ahead by 4.9 |
Microsoft’s footnote says other models’ numbers come from their official model cards, and a dash means no reported score. GPT 5.4 has no score on the other five benchmarks, so no comparison is possible there.
The table has no GPT-5.6 column, although Microsoft’s July 23 post compares an Excel model with GPT-5.6.
Against Anthropic’s models Microsoft’s claims are stronger: 52.8 on SWE-Bench Pro versus 53.4 for Claude Opus 4.6, and human raters preferring it to Claude Sonnet 4.6 across 1,276 tasks. The release reports that human test only against Sonnet 4.6. All of these are Microsoft’s own results.
The release describes a sparse mixture-of-experts model with 35 billion active and about 1 trillion total parameters, a 256k-token context window, function calling and Chat Completions API compatibility.
Where Microsoft says its own models already stand in for OpenAI’s
| Surface | Microsoft’s claim | Limit |
|---|---|---|
| MAI-Thinking-1 on Foundry | Public preview since August 12 | The release names no Microsoft product that uses it |
| GitHub Copilot (MAI-Code-1-Flash) | About 10% higher code accept rate and 10% lower median token use than GPT 5.4 Mini and Claude Haiku 4.5 in VS Code (July 23) | Compared with small models, and a different model from MAI-Thinking-1 |
| Excel (MAI model tuned from MAI-Code-1-Flash) | Quality “on par with GPT-5.6 for the most common tasks” at lower cost (July 23) | Based on user feedback from production traffic; common tasks only |
| Copilot Chat, Outlook, PowerPoint | Microsoft is “currently extending” the approach (July 23) | A stated plan; no results in that post |
Which Microsoft 365 surfaces use MAI models, and where GPT-5.6 is the preferred model, is covered in our earlier post. Three decision rules follow from the numbers above:
- Agentic terminal work. Microsoft’s table shows a 29.1-point gap to GPT 5.4 on Terminal Bench 2.0. Test your own tasks before relying on MAI-Thinking-1.
- Competition-style math or LiveCodeBench. Microsoft reports no GPT 5.4 score there, so its table cannot tell you which model is better.
- Spreadsheet or code-completion workloads. Microsoft’s published results come from other, smaller MAI models, not MAI-Thinking-1.
How Much Does Microsoft Still Depend on OpenAI?
Microsoft’s FY2026 10-K says it recorded $24.1 billion of revenue from commercial arrangements with OpenAI, inclusive of revenue-sharing payments, with $6.0 billion receivable at June 30, 2026. Total FY2026 revenue was $331,839 million, so the OpenAI line is 24,100 / 331,839 = 7.3%.
The 10-K gives no AI-revenue total, so the often-cited 70% is not a Microsoft figure. Bloomberg estimated that about 70% of Microsoft’s AI sales came from OpenAI by assuming the 123% growth rate Microsoft reported for March continued, implying about $34 billion of AI revenue (24.1 / 34 = 71%).
The $24.1 billion also mixes payment types: OpenAI pays Microsoft for cloud computing and model-building costs as well as a share of its revenue (The Next Web). It measures the relationship’s size, not how much a Microsoft-built model could replace.
Two Lenses
Building the exit before you have to use it
Microsoft’s own language for MAI-Thinking-1 is specific: the model was trained “from the ground up on clean, traceable and enterprise-grade data, without distillation from third-party models.” The release does not name OpenAI or say the model is meant to replace anything.
The release also lists “self-sufficiency across the entire stack” as a pillar, from “co-design of our models with MSFT’s own accelerators” to in-house training infrastructure. Reading one: Microsoft is building the capacity to need OpenAI less.
Microsoft’s June 2 launch post adds that co-designing with its own Maia 200 silicon is “already” producing “a 1.4x efficiency boost,” which it calls “long term self-sufficiency for Microsoft and our partners.”
Cashing the check while you build the exit
Reading two: the commercial relationship continues. Microsoft says OpenAI’s revenue-share payments continue through 2030 at the same percentage, subject to a total cap. The cap’s level is not published, so its dollar effect cannot be computed from public documents.
Azure also kept a privileged position: OpenAI’s products still “ship first on Azure, unless Microsoft cannot and chooses not to support the necessary capabilities,” and Microsoft “remains OpenAI’s primary cloud partner” (Microsoft, April 27).
The 10-K figures above show what the continuing relationship was worth in FY2026: $24.1 billion. Building an alternative supplier is a hedge; Microsoft’s own table does not yet show a replacement for GPT 5.4.
Does Giving Up Exclusivity Cost Microsoft Money?
Not in a way public documents let us measure. The April change made Microsoft’s IP license non-exclusive, and payments continue through 2030 at the same percentage with a total cap. The cap’s amount is not public, so any dollar effect is unknown.
OpenAI can now serve its products on any cloud, including Amazon and Google (CNBC, April 27). No public document we reviewed quantifies whether that shifts workloads away from Azure.
Three questions sort a future Microsoft-OpenAI headline from noise:
- Does it touch the 2032 IP license, or only where training data comes from? Those are different questions with different answers today.
- Does money move between the two companies, or only get capped? The April 2026 change added a total cap; its level is not public.
- Does a new Microsoft model replace OpenAI in a shipping product, or run alongside it? MAI-Thinking-1 is in Foundry public preview; the product results Microsoft has published come from other MAI models.
What would change our view
If Microsoft named MAI-Thinking-1 as the default model in a flagship Copilot product, that would be a materially stronger signal than a preview in a developer catalog.
A new Microsoft benchmark table that scores MAI-Thinking-1 against GPT-5.6 on the same tests would show whether the gaps above are closing.
FAQ
Q. Is Microsoft still OpenAI’s exclusive cloud provider?
A. No. As of April 27, 2026, OpenAI can serve products through any cloud provider. Microsoft kept a narrower right: OpenAI’s products still ship first on Azure unless Microsoft opts out of supporting them.
Q. Does Microsoft still make money from OpenAI?
A. Yes. OpenAI continues paying a revenue share through 2030 at the same percentage, subject to a total cap. Microsoft’s FY2026 10-K shows $24.1 billion of revenue from commercial arrangements with OpenAI, about 7.3% of total revenue; Bloomberg estimates about 70% of Microsoft’s AI sales.
Q. Is MAI-Thinking-1 meant to replace OpenAI’s models inside Microsoft’s products?
A. Not on the evidence Microsoft has published. In its own benchmark table MAI-Thinking-1 trails GPT 5.4 on all three benchmarks where both have scores, and its published product results come from other MAI models.
Q. How does MAI-Thinking-1 compare with GPT 5.4?
A. In Microsoft’s table: GPQA Diamond 84.2 vs 92.8, Terminal Bench 2.0 46.0 vs 75.1, SWE-Bench Pro 52.8 vs 57.7. GPT 5.4 has no score on the other five benchmarks, so they cannot be compared.
Sources
- The next chapter of the Microsoft-OpenAI partnership — Microsoft, Oct 28, 2025
- Joint statement on continuing partnership — Microsoft, Feb 27, 2026
- The next phase of the Microsoft-OpenAI partnership — Microsoft, Apr 27, 2026
- Introducing MAI-Thinking-1 — Microsoft AI, Aug 12, 2026
- Microsoft’s AI sales mostly come from OpenAI — Bloomberg, Aug 5, 2026
- OpenAI shakes up partnership with Microsoft, capping revenue share — CNBC, Apr 27, 2026
- Building a hill-climbing machine: Launching seven new MAI models — Microsoft AI, June 2, 2026
- Hill-climbing MAI models for GitHub Copilot and Excel — Microsoft AI, July 23, 2026
- Microsoft Form 10-K for fiscal year ended June 30, 2026 — SEC EDGAR
- Microsoft’s AI revenue is ~70% OpenAI, a filing shows — The Next Web, August 2026
Correction, October 2, 2026. An earlier version said Microsoft shipped MAI-Thinking-1 on June 2, 2026 and took eight months to leave OpenAI exclusivity.
Microsoft announced the model on June 2 and opened a public preview on August 12, and the license changed from exclusive to non-exclusive between October 2025 and April 2026.
Update, October 2, 2026. This post was reworked around MAI-Thinking-1 and OpenAI dependence, with the contract timeline moved to our explainer. An earlier version attributed the 70% figure to Microsoft’s disclosures; it is Bloomberg’s estimate. Unsourced claims about the Maia chip, a Microsoft executive quote and negotiation length were removed.
Leave a Reply