Microsoft just ended its exclusive reliance on OpenAI. At Build 2026, the company unveiled seven in-house AI models trained entirely from scratch -- zero OpenAI code, data, or distillation anywhere in the pipeline. The lineup spans reasoning, coding, image generation, voice, and transcription, marking Microsoft's first full-stack frontier AI portfolio.
For developers and enterprises watching the AI cost curve, this changes the vendor landscape overnight. GitHub Copilot now offers MAI-Code-1-Flash as a selectable model alongside OpenAI and Anthropic options, priced at $0.75 per million input tokens and $4.50 per million output tokens. Microsoft claims it beats Claude Haiku 4.5 by 16 percentage points on SWE-Bench Pro while using 60 percent fewer tokens on complex tasks.
Why Microsoft Built Its Own Models Now
The timing is deliberate. OpenAI and Microsoft formally ended their exclusive partnership and revenue-sharing agreement on April 27, 2026, according to Forbes reporting. That contract had locked Microsoft into OpenAI as its sole frontier model supplier since the original $1 billion investment in 2019, followed by roughly $2 billion more through 2023 and a $10 billion multiyear deal in January 2023.
After the October 2025 restructuring where OpenAI became a public benefit corporation, Microsoft held roughly 27 percent of OpenAI Group PBC -- a stake valued near $135 billion -- while OpenAI committed to $250 billion in incremental Azure services. But the April 2026 exclusivity termination freed both sides to build and sell competing products. Microsoft's MAI models are the first tangible result.
This is not a breakup. Microsoft keeps its equity stake and OpenAI keeps running on Azure. But the relationship shifted from sole supplier to one option among several, with Microsoft's own models now sitting on the shelf next to OpenAI's.
MAI-Thinking-1: Microsoft's First Independent Reasoning Model
MAI-Thinking-1 is a 35-billion active-parameter sparse Mixture-of-Experts model with a 256,000-token context window. Microsoft reports 97.0 percent on AIME 2025 and parity with Claude Opus 4.6 on coding benchmarks. In blind testing by independent raters at Surge, MAI-Thinking-1 was reportedly preferred over Claude Sonnet 4.6 across both single-turn and multi-turn tasks.
The model entered private preview on Microsoft Foundry on June 2, 2026. Enterprises can request access, but Microsoft has not committed to a general-availability date. This cautious rollout contrasts with the coding model's immediate Copilot integration.
MAI-Code-1-Flash: The Developer-Facing Model
MAI-Code-1-Flash is a sparse Mixture-of-Experts system with 5 billion active parameters out of 137 billion total, also with a 256K context window. It began rolling out June 2 across every GitHub Copilot tier: Free, Pro, Pro+, and Max, accessible through the VS Code model picker.
Microsoft claims it outperforms Claude Haiku 4.5 by 16 points on SWE-Bench Pro while using 60 percent fewer tokens on complex tasks. GitHub lists pricing at $0.75 per million input tokens and $4.50 per million output tokens -- a budget option tuned for VS Code and GitHub Copilot CLI workflows rather than general-purpose chat.
At scale, the token efficiency compounds. A team pushing 40 million input and 8 million output tokens monthly would spend roughly $66 on MAI-Code-1-Flash versus $240 on GPT-5.6 Sol at blended rates -- a 72 percent reduction. These are illustrative figures using public per-token pricing, not official benchmarks, but they explain why finance teams are tracking the rollout.
The Full MAI Portfolio: Seven Models, One Strategy
Beyond the two flagships, Microsoft announced five additional models: MAI-Image-2.5, MAI-Image-2.5 Flash, MAI-Transcribe-1.5, MAI-Voice-2, and MAI-Voice-2 Flash. The naming convention (full and Flash variants) mirrors Google's Gemini strategy and Microsoft's own reasoning/coding pair -- a deliberate playbook pairing a slower, higher-quality model for accuracy-sensitive work with a faster, cheaper variant for high-volume tasks.
| Model | Category | Active Parameters | Pricing | Availability |
|---|---|---|---|---|
| MAI-Thinking-1 | Reasoning | 35B (MoE) | Not disclosed | Private preview, Microsoft Foundry |
| MAI-Code-1-Flash | Coding | 5B / 137B total (MoE) | $0.75 in / $4.50 out | GitHub Copilot, all tiers |
| MAI-Image-2.5 | Image generation | Not disclosed | Not disclosed | Announced, limited rollout |
| MAI-Image-2.5 Flash | Image generation (fast) | Not disclosed | Not disclosed | Announced, limited rollout |
| MAI-Transcribe-1.5 | Transcription | Not disclosed | Not disclosed | Announced, limited rollout |
| MAI-Voice-2 | Voice/speech | Not disclosed | Not disclosed | Announced, limited rollout |
| MAI-Voice-2 Flash | Voice/speech (fast) | Not disclosed | Not disclosed | Announced, limited rollout |
Microsoft has stated internal deployments for clients like McKinsey delivered performance surpassing GPT-5 at roughly ten times the cost efficiency -- a vendor claim not yet independently reproduced.
How MAI Models Stack Up Against the Field
No single benchmark captures cross-vendor comparison since Microsoft, OpenAI, Anthropic, Google, and Chinese labs report on different tests. Here is how current flagships line up on published numbers:
| Model | Maker | Type | Context | Headline Benchmark | Status (June 2026) |
|---|---|---|---|---|---|
| MAI-Thinking-1 | Microsoft | Reasoning | 256K tokens | 97.0% AIME 2025 | Private preview |
| MAI-Code-1-Flash | Microsoft | Coding | 256K tokens | +16 pts vs Haiku 4.5 SWE-Bench | Live in Copilot |
| Claude Opus 4.8 | Anthropic | General/Coding | Not disclosed | 69% SWE-Bench | Generally available |
| GPT-5.6 Sol | OpenAI | General/Coding | Not disclosed | 88.8% TerminalBench | Generally available |
| Gemini 3.5 Flash | General | Not disclosed | 4x faster than prior Flash | Default AI in Search | |
| GLM-5.2 | Zhipu AI | General/Coding | Not disclosed | Beats GPT-5.5 at 1/6 cost | Open-weight, GA |
The pattern is consistent: no vendor wins on both price and performance simultaneously. Microsoft's MAI models target that gap for Azure and GitHub customers.
What This Means for Enterprise Buyers
The commercial logic is straightforward. Every GitHub Copilot request routing to an OpenAI model carried third-party inference costs. Routing meaningful share to MAI-Code-1-Flash means Microsoft captures more margin on infrastructure it already owns. Microsoft's Build 2026 recap frames this explicitly around cost control and reduced outside dependence.
For OpenAI, the effect is nuanced. The $250 billion Azure compute commitment remains, so Microsoft's MAI success does not necessarily hurt OpenAI's infrastructure revenue. But OpenAI's models now compete for Copilot traffic against a Microsoft-made alternative that Microsoft has every incentive to route users toward by default. Expect pricing pressure across the mid-tier coding segment where MAI-Code-1-Flash, Claude Haiku 4.5, and Chinese open-weight models all fight for budget-conscious developers.
Canadian enterprises got a timely signal. Two days after the MAI launch, Prime Minister Mark Carney launched "AI for All" -- Canada's national AI strategy targeting $200 billion in economic growth, 250,000 new AI jobs, and lifting business AI adoption from 12 percent to 60 percent by 2034. The strategy flags concern about "critical parts of our AI ecosystem" falling under foreign control. For Canadian banks, hospitals, or agencies already weighing data residency and vendor concentration, an Azure-native model stack inside existing Microsoft 365 and GitHub Enterprise contracts is lower friction than onboarding a new vendor -- though not truly sovereign since MAI models are still built and hosted by a US company.
Risks and Open Questions
Several things remain unresolved. First, Microsoft's benchmark numbers are self-reported. The Surge blind evaluation is a useful third-party signal but not an independent lab replication. Second, MAI-Thinking-1's private-preview status means most claims cannot yet be tested by outside developers. Third, safety guardrails on closed-weight, API-gated models rely on Microsoft's internal testing rather than public weight auditing -- a trade-off versus open-weight alternatives where protections have been stripped in minutes by motivated users.
5 Predictions for the Rest of 2026
-
MAI models expand beyond Copilot. Expect MAI-Thinking-1 and MAI-Code-1-Flash in Microsoft 365 Copilot and Azure AI Foundry before year-end.
-
OpenAI responds on price. With Azure enterprise customers able to default to a Microsoft-made alternative, look for OpenAI to sharpen enterprise discounting.
-
Independent benchmarking catches up. Third-party labs should publish SWE-Bench Pro and AIME reruns on MAI models by midsummer.
-
Canadian RFPs ask for "OpenAI-optional" architectures. Sovereignty framing plus Azure-native alternatives will drive procurement language avoiding single upstream providers.
-
Other hyperscalers accelerate in-house bets. Amazon and Google already have internal model lines; pressure to reduce external lab reliance will grow through H2 2026.
Frequently Asked Questions
What is Microsoft MAI-Thinking-1?
MAI-Thinking-1 is Microsoft's first in-house reasoning model with 35 billion active parameters on a sparse Mixture-of-Experts architecture and a 256,000-token context window. It entered private preview on Microsoft Foundry on June 2, 2026.
What is MAI-Code-1-Flash and how much does it cost?
MAI-Code-1-Flash is Microsoft's first in-house coding model with 5 billion active parameters out of 137 billion total. GitHub lists pricing at $0.75 per million input tokens and $4.50 per million output tokens, live across all GitHub Copilot tiers.
Does GitHub Copilot now use Microsoft's own AI models instead of OpenAI's?
Copilot now offers MAI-Code-1-Flash as a selectable option in the VS Code model picker alongside existing models. It is an added choice, not a full replacement of OpenAI or Anthropic models within Copilot.
Why is Microsoft reducing its reliance on OpenAI now?
OpenAI and Microsoft formally ended their exclusive partnership and revenue-sharing agreement on April 27, 2026. That change freed both companies to build and sell competing products, and MAI models are Microsoft's first major result.
Are Microsoft's MAI models available in Canada?
MAI-Code-1-Flash is rolling out globally across GitHub Copilot tiers including Canadian accounts. MAI-Thinking-1 remains in private preview on Microsoft Foundry with request-based access.
How does MAI-Code-1-Flash compare to Claude Haiku 4.5?
Microsoft claims MAI-Code-1-Flash outperforms Claude Haiku 4.5 by 16 percentage points on SWE-Bench Pro while using 60 percent fewer tokens on complex coding tasks. This figure comes from Microsoft's launch materials and has not been independently reproduced.
Is Microsoft ending its partnership with OpenAI entirely?
No. Microsoft still holds roughly 27 percent stake in OpenAI Group PBC valued near $135 billion, and OpenAI remains committed to $250 billion in incremental Azure services. What ended was exclusivity and revenue sharing, not the investment relationship.
What are the other five MAI models Microsoft announced?
Alongside MAI-Thinking-1 and MAI-Code-1-Flash, Microsoft unveiled MAI-Image-2.5, MAI-Image-2.5 Flash, MAI-Transcribe-1.5, MAI-Voice-2, and MAI-Voice-2 Flash covering image generation, voice, and transcription. Detailed specs and pricing for these five had not been published as of Build 2026.
Key Takeaways
- Microsoft launched seven in-house MAI models trained from scratch with zero OpenAI dependency
- MAI-Code-1-Flash is live in GitHub Copilot at $0.75/$4.50 per million tokens, claiming 16-point SWE-Bench advantage over Claude Haiku 4.5
- The April 2026 exclusivity termination with OpenAI enabled this move; equity and Azure commitments remain
- Enterprises gain an Azure-native alternative for vendor diversification and cost control
- Independent benchmarking and broader rollout across Microsoft 365 and Azure AI Foundry are the next signals to watch
Conclusion
Microsoft's MAI models represent a structural shift in the AI vendor landscape. The company that spent $13 billion funding OpenAI now ships its own frontier models across reasoning, coding, image, voice, and transcription. For developers, this means more choice and competitive pricing inside existing toolchains. For enterprises, it means vendor diversification without new procurement cycles. The real test comes when third-party benchmarks validate or challenge Microsoft's claims -- and whether MAI-Thinking-1 graduates from private preview to general availability before 2027.
Sources: Tech Insider Canada - Microsoft's 7 MAI Models Cut OpenAI Dependence, Forbes - OpenAI and Microsoft End Exclusive Partnership, Microsoft AI Blog - Building a Hill-Climbing Machine, Microsoft AI Blog - Introducing MAI-Code-1-Flash, Microsoft Build 2026 Recap, Prime Minister of Canada - AI for All Strategy
Automated Transmission
This entry was synthesized and populated dynamically using native API integrations.