Question 1

What models did Microsoft announce at Build 2026?

Accepted Answer

Microsoft announced seven MAI (Microsoft AI) models at Build 2026: MAI-Thinking-1 for extended reasoning, MAI-Code-1 for full-precision code generation, MAI-Code-1-Flash for low-latency inline completions, MAI-DS-R1 for data science and structured analytics, MAI-Vision-1 for multimodal document understanding, MAI-Mini as a sub-3B parameter edge model for on-device deployment, and MAI-Embed-1 for embeddings and semantic retrieval. MAI-Code-1-Flash launched as the default completion model in GitHub Copilot at Build, replacing the previous OpenAI Codex and GPT-4o completions. MAI-Code-1, MAI-Mini, and MAI-Embed-1 are generally available on Azure; MAI-Thinking-1, MAI-DS-R1, and MAI-Vision-1 are in preview.

Question 2

How does MAI-Thinking-1 compare to Claude 3.7 Sonnet and GPT-4.5?

Accepted Answer

On mathematical and scientific reasoning benchmarks, MAI-Thinking-1 is competitive at the frontier: 82.4% on MATH competition problems versus Claude 3.7 Sonnet's 81.2% and GPT-4.5's 80.8%, and 68.1% on GPQA Diamond expert reasoning versus Claude's 71.0% and GPT-4.5's 67.3%. The gap is most pronounced on software engineering: MAI-Code-1 scores 61.2% on SWE-bench Verified while Claude 3.7 Sonnet with extended thinking scores 70.3%. For enterprise teams using AI in code review, refactoring, and software development workflows, Anthropic retains a meaningful capability advantage. For scientific analysis, financial modeling, and research summarization, MAI-Thinking-1 benchmarks within margin of error of the frontier models at a pricing tier comparable to Claude 3.5 Sonnet on Azure.

Question 3

What does 'no OpenAI data' mean for enterprise AI compliance?

Accepted Answer

Microsoft's MAI models were trained without OpenAI-licensed datasets, synthetic data derived from ChatGPT outputs, or content from OpenAI training data pools. This has practical implications for enterprises operating under EU AI Act requirements, which mandate training data provenance documentation for General-Purpose AI models. Regulated industries — financial services, healthcare, and legal services — increasingly include AI training data provenance in vendor qualification requirements. Microsoft's clean provenance architecture means enterprise legal and compliance teams can obtain documentation showing exactly what licensed data trained the MAI models, without exposure to third-party intellectual property licensing entanglements associated with models that used GPT-4 outputs during training. This is a meaningful differentiator for procurement in regulated verticals.

Question 4

Does the MAI launch change Microsoft's OpenAI partnership?

Accepted Answer

Microsoft has framed the MAI launch as model diversity rather than a departure from OpenAI. OpenAI models remain available through Azure OpenAI Service, and Microsoft describes the partnership as intact. What changes materially is Microsoft's negotiating leverage: it now enters partnership renegotiations with a credible in-house alternative that has demonstrated production readiness in GitHub Copilot. The financial impact on OpenAI — which distributes a substantial share of API revenue through Azure — depends on how much workload Microsoft shifts to MAI models over the next 12 months. Enterprise customers evaluating the MAI announcement should also assess how multi-model routing through Agent 365 changes their own negotiating leverage with both Microsoft and OpenAI on upcoming contract renewals.

Question 5

What is Project Solara and when did Microsoft start it?

Accepted Answer

Project Solara is Microsoft's internal AI model research program, reportedly operational since approximately 2022. It employs more than 400 researchers and engineers in a track separate from the Azure OpenAI Service team, with dedicated compute clusters, training infrastructure, and data sourcing agreements independent of the commercial OpenAI partnership. The program's early outputs include Microsoft's Phi series of small language models, released openly starting in 2023. Solara's operational independence from the OpenAI integration team was intentional: it ensures Microsoft's in-house AI roadmap can proceed regardless of partnership terms. The MAI family announced at Build 2026 is the program's first major generally available product output, with MAI-Code-1-Flash already deployed to GitHub Copilot's global subscriber base.