Microsoft’s first-party AI stack just crossed a threshold: every layer is now either in production or public preview. The MAI family — trained from scratch on Microsoft’s own Maia 200 silicon, with zero distillation from OpenAI, Anthropic, or any third-party lab — has moved from Build 2026 announcements to shipping reality in under three months. For CTOs mapping vendor dependencies, the question is no longer “does Microsoft have its own models?” but “how good are they, and what do I standardize on?”

This week’s edition covers MAI-Thinking-1’s public preview, the economics of MAI-Code-1.1-Flash and MAI-Image-2.6, two hard migration deadlines, and the confirmed Phi Silica to Aion transition.

MAI-Thinking-1: First Fully In-House Frontier Reasoning Model

On August 12, MAI-Thinking-1 entered public preview on Microsoft Foundry — Microsoft’s first frontier reasoning model trained entirely from scratch. Architecture: sparse mixture-of-experts, roughly 1 trillion total parameters with 35 billion active, and a 256K context window.

Self-reported numbers are strong: 97.0% on AIME 2025, 94.5% on AIME 2026, and 52.8% on SWE-Bench Pro, matching Claude Opus 4.6. Microsoft also cites a blind preference win over Claude Sonnet 4.6 across 1,276 tasks.

The honest caveat for engineering teams: independent benchmarks are still scarce. Early third-party analysis from Bloomberg and ByteIota places MAI-Thinking-1 roughly equivalent to DeepSeek V3.2 in real-world use. That is not a dismissal — it is context. The point of public preview is that you can now run your own eval suite against it on Foundry instead of trusting vendor numbers.

MAI-Code-1.1-Flash: The Economics of First-Party Code

MAI-Code-1.1-Flash went GA on August 11 and is now the default coding model across every GitHub Copilot tier. The headline is cost: $0.20/1M input and $1.20/1M output — a 75% cut from the original MAI-Code-1-Flash — with 25% faster token streaming and 25% fewer tokens per task.

Capability gains matter as much as price. Native image input arrives for the first time in the MAI-Code line, enabling screenshot-to-code workflows. SWE-Bench Verified is up to 72.6%, Terminal-Bench 2.1 jumped 22% to 62.9%, and .NET tasks improved 15%. The internal ScreenShot2WebApp score of 42.1% shows how early the vision-coding path still is — but the direction is clear.

Deprecation watch: MAI-Code-1-Flash retires September 10. If you pinned prompts or evals to it, migrate this week.

MAI-Image-2.6: #2 on the Arena Leaderboard

MAI-Image-2.6 launched August 10 and immediately hit #2 on the Arena text-to-image leaderboard, ahead of Google, Meta, and xAI — behind only OpenAI. It ranks #3 on image editing. The +79 Elo gain over 2.5 overall and +91 on text rendering reflect real quality jumps in typography, portraits, and photorealistic commercial output.

The number CTOs should care about most: Microsoft reports an 84% GPU cost reduction versus GPT-Image-2 in production. That is the Maia 200 co-design story — silicon and models built together — and it is the same pattern behind the MAI-Code price cut. First-party economics are becoming a structural advantage, not a marketing claim.

Purpose-Built Security: MAI-Cyber-1-Flash

MAI-Cyber-1-Flash, in public preview since August 3 via Project Perception, is Microsoft’s entry into purpose-built security AI — vulnerability detection from a model trained for the task rather than a general LLM pointed at it. For security teams, this is the first-party answer to “fine-tune everything”: a dedicated model with security-specific training, routed through Copilot’s orchestration layer.

Voice and Transcription

MAI-Voice-2 is GA with 15 languages and short-sample voice cloning; MAI-Voice-2-Flash offers 2× speed at 32% lower cost in public preview. MAI-Transcribe-1.5 handles 43 languages at a claimed 5× competitor speed. Notably, a full-duplex MAI Realtime voice model has been spotted in the MAI Playground but is not yet officially announced — a signal of what is coming.

Phi-4: The Open-Weight Edge Portfolio

The Phi-4 family remains the industry’s most comprehensive open-weight SLM portfolio: seven MIT-licensed variants from 3.8B to 15B parameters. The newest member, Phi-4-reasoning-vision-15B (March 2026), pairs a SigLIP-2 vision encoder with the reasoning backbone, supports screen grounding, and toggles reasoning on demand — the best on-premise vision-reasoning candidate at its size.

Two numbers keep coming up in client conversations: Phi-4 (14B) matches Llama 3.3 70B-class MMLU at 84.8% with 4–5× less VRAM, and Phi-4-mini (3.8B) sustains 1,955 tokens/sec on Intel Xeon 6 in CPU-only mode. For data-sovereign or air-gapped deployments, there is still no cheaper on-ramp to capable AI than MIT-licensed Phi weights.

Florence-2: 13 Days to the Hard Cutover

Florence-2 remains the vision engine behind Azure AI Vision Image Analysis 4.0, handling 12+ vision tasks from a single weight set, including OCR across 164 languages. Microsoft has confirmed no Florence-3 is planned — stability over release velocity.

The deadline that matters: legacy Azure Vision APIs (v1.0–3.1) retire hard on September 13 — 13 days out. If any workload still calls the old APIs, migrating to the Florence-2-based Image Analysis 4.0 SDK is the highest-priority item on this list.

Turing: Effectively Retired

No meaningful updates in 2025–2026. MAI has taken over every production workload Turing served. If Turing models still appear in your architecture diagrams, those diagrams are now historical documents — plan the replacement.

Aion 1.0: The On-Device Transition Is Firm

The Phi Silica to Aion transition now has hard dates: standalone Aion Instruct testing October 1, Windows Insider rollout October 23, GA November 24 — at which point Phi Silica is removed from retail devices.

Aion 1.0 Instruct runs on CPU, GPU, or NPU — no dedicated AI hardware required — with open weights expected and no Limited Access Feature token requirement. Aion 1.0 Plan (14B, 32K context) targets on-device agentic workflows: reasoning, tool-calling, file management, sub-agent orchestration. One migration gotcha: existing Phi Silica LoRAs will need retraining for Aion.

Copilot: A First-Party Orchestration Layer

Copilot remains a multi-vendor orchestrator — a task-categorization layer routes each interaction to the best model. MAI handles email drafting, summarization, spreadsheets, coding, image generation, transcription, and security scanning. GPT-5.6 and Claude stay in the mix for complex multi-step reasoning and creative work. Microsoft’s projection: by end of 2026, most commodity Copilot tasks run entirely on MAI. Plan cost and data-residency models around that trajectory.

Action Items

  1. Migrate MAI-Code-1-Flash to 1.1-Flash before September 10 — 75% cost cut plus image input
  2. Migrate legacy Azure Vision APIs to Image Analysis 4.0 before September 13 — hard retirement
  3. Benchmark MAI-Thinking-1 yourself on Foundry — independent numbers beat vendor claims
  4. Evaluate MAI-Image-2.6 economics if you generate images at scale — 84% GPU cost reduction vs GPT-Image-2
  5. Audit for Turing dependencies and plan Phi Silica LoRA retraining ahead of the November 24 Aion cutover

Kevin Kaminski is the founder of Big Hat Group Inc., a Microsoft partner focused on AI adoption strategy. Follow at https://x.com/kkaminsk for ongoing analysis of Microsoft’s AI model ecosystem.