Microsoft’s first-party AI model flywheel is now self-reinforcing. This week’s evidence: MAI-Transcribe-2 claims the top spot on FLEURS across 60 languages at 5.2% average word error rate, MAI-Image-2.6 debuts at No. 2 on the Arena text-to-image leaderboard, and MAI-Code-1.1-Flash is now the default GitHub Copilot coding model — with a 75% cost cut over its predecessor. Meanwhile, the legacy Azure Vision APIs that powered a generation of vision pipelines went 410 Gone on September 13.

For engineering leaders, the strategic picture is crystallizing. Microsoft is no longer shipping first-party models as alternatives to third-party options — it is routing production traffic to them by default, pricing them aggressively, and retiring the infrastructure that competing paths depend on. This edition covers the three MAI releases that landed in the last two weeks, the Copilot first-party routing shift, the Azure Vision API retirement fallout, and the Aion transition timeline that Windows teams must plan for now.

MAI-Transcribe-2: Fastest, Cheapest, Most Accurate Speech Recognition

MAI-Transcribe-2 entered public preview on September 3, 2026, and Microsoft’s claims are bold: No. 1 on FLEURS across 60 languages at 5.2% average word error rate, No. 2 on Artificial Analysis AA-WER at 2.0%, and roughly 10× the speed of GPT-Transcribe, 7× Scribe v2, and 5× Gemini 3.5 Transcribe. The pricing is aggressive: $0.10 per hour of audio transcribed, a limited-time rate running through end of 2026.

The model supports diarization and word-level timestamps natively — no post-processing pipeline required. For teams running large-scale transcription workloads (call center analytics, meeting capture, media indexing), the economics are significant. At $0.10/hour, transcribing 10,000 hours of audio costs $1,000 — a fraction of what comparable cloud speech APIs charge. The 60-language coverage with a single model eliminates the multi-model routing complexity that plagues existing pipelines.

The caveat: these are Microsoft-reported benchmarks. Independent validation will follow as broader API access opens. But the pricing is verifiable on Microsoft Foundry today, and the performance claims are consistent with the trajectory of MAI-Transcribe-1 and 1.5.

MAI-Image-2.6: Arena No. 2 With 84% GPU Cost Reduction

MAI-Image-2.6 launched on August 10, 2026, with the full reveal on September 4. The headline numbers: No. 2 on the Arena text-to-image leaderboard (+79 Elo over version 2.5) and No. 1 on Artificial Analysis for image editing. The Flash variant is 2.8× faster and 72% more efficient than GPT-Image-2-Medium.

The cost story is where this gets interesting for production deployments. Microsoft reports an 84% GPU cost reduction versus GPT-Image-2 in production workloads. The model supports multi-image reference editing, web grounding for fact-checked image generation, and dynamic aspect ratios — features that previously required separate pipelines.

Pricing on Microsoft Foundry: MAI-Image-2.5 tiers at $19.50 (Flash), $47 (standard), and $106 (Pro) per million image outputs. For organizations generating images at scale — marketing teams, e-commerce catalogs, content production — the GPU cost reduction translates directly to infrastructure spend. The image-editing capability climbing from No. 5 to No. 3 on Arena (+19 Elo) means this is not just a generation model; it is a credible editing tool.

MAI-Code-1.1-Flash: 75% Cost Cut, Now Default Across Copilot

MAI-Code-1.1-Flash went generally available on August 11, 2026, and immediately became the default GitHub Copilot coding model across all tiers — Free through Max. The pricing: $0.20 per million input tokens and $1.20 per million output tokens, a 75% cost reduction from the original MAI-Code-1-Flash. Token streaming is 25% faster, token consumption per task is 25% lower, and the model now accepts image inputs for diagram-to-code scenarios.

The previous MAI-Code-1-Flash was retired on September 10, 2026. Teams with pinned prompts or evaluation harnesses targeting that model must migrate to 1.1-Flash. GPT-4 Turbo remains available as a fallback through November 2026, after which MAI-Code-1.1-Flash runs exclusively on Microsoft’s Maia 200 accelerators.

The routing reality: Bloomberg confirmed that Microsoft 365 Copilot now routes Excel and Outlook prompts to MAI models in production. Copilot is not a single model — it is an orchestration layer with a task categorization router. Email drafting, summarization, Excel formula generation, and structured data extraction route to MAI. Agentic coding routes to MAI-Code-1.1-Flash. Image generation routes to MAI-Image-2.5 and 2.6. Teams transcription routes to MAI-Transcribe. Security scanning routes to MAI-Cyber-1-Flash. Only complex multi-step reasoning and creative generation requiring stylistic judgment still routes to third-party frontier models.

The implication for architecture planning: budget on the assumption that first-party routing expands, not contracts. Microsoft projects that most commodity Copilot tasks will run on MAI by end of 2026.

Azure Vision API Retirement: 410 Gone As Of Yesterday

September 13, 2026 was the hard retirement date for legacy Azure Vision APIs versions 1.0 through 3.1. Any application still calling these endpoints now receives 410 Gone responses. This is not a soft deprecation — there is no gradual ramp-down, no extended grace period.

The migration target is Florence-2-based Image Analysis 4.0 SDK. Florence-2 remains Microsoft’s vision foundation model: a unified sequence-to-sequence architecture handling 12+ vision tasks with a single weight set — detection, segmentation, captioning, dense region captioning, visual grounding, and OCR across 164 languages. The large variant (771M parameters) is production-grade; the base variant (232M) runs on standard CPUs. Both are MIT-licensed on Hugging Face with roughly 2.66 million and 813,000 monthly downloads respectively.

No Florence-3 is planned. Microsoft is prioritizing stability over release velocity for the vision foundation layer. Image Analysis 4.0 itself retires September 25, 2028 — giving teams a two-year horizon to plan around.

If you have not yet migrated, this is a production-impacting outage. Audit your applications for any calls to the old vision endpoints today.

Aion 1.0: The Phi Silica Cutover Timeline

Aion 1.0 is Microsoft’s on-device replacement for Phi Silica, announced at Build 2026. Two variants: Aion 1.0 Instruct (lightweight SLM for summarization, rewriting, intent classification — runs on CPU, GPU, or NPU with no dedicated hardware) and Aion 1.0 Plan (14B parameters, 32K context, for on-device agentic reasoning via the Windows Agent Framework).

The transition has hard dates:

  • October 1, 2026: Standalone Aion Instruct testing package with LoRA training support.
  • October 23, 2026: Windows Insiders rollout — both Phi Silica and Aion coexist, with a Controlled Feature Rollout selecting the active model per device.
  • November 24, 2026: General availability — Aion Instruct becomes the production model and Phi Silica is removed from retail devices.

The critical planning item: existing Phi Silica LoRA adapters will not transfer to Aion. If your team has invested in custom Phi Silica fine-tunes, begin retraining now. API compatibility is preserved across the LanguageModel and ImageDescription APIs, but the underlying model weights are different. Final GPU compatibility lists and required NVIDIA driver versions are not yet published — watch for those before the October dates.

Phi-4: The Open-Weight Workhorse Continues

While MAI dominates the cloud frontier, Phi-4 remains Microsoft’s open-weight small language model family for edge and IoT. The latest release, Phi-4-reasoning-vision-15B (March 2026), combines the Phi-4-reasoning backbone with a SigLIP-2 vision encoder for multimodal reasoning — 15B parameters, MIT license, dynamic thinking activation. Phi-4-mini shipped to Windows 12 Copilot+ PCs in July 2026, delivering 42 tokens per second on Snapdragon X Elite at roughly 3GB VRAM for offline code completion and local image editing.

The Phi family’s punch-above-weight story continues: Phi-4 (14B) outperforms models 5× its size on math and reasoning benchmarks, driven by synthetic data quality and refined training curricula rather than parameter count. For teams needing on-prem or air-gapped deployments, Phi-4 remains the benchmark.

Turing: Treat As Retired

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

What to Track Next

  • October 1: Aion 1.0 Instruct standalone testing package.
  • October 23: Aion 1.0 Instruct reaches Windows Insiders.
  • November 24: Aion 1.0 GA; Phi Silica removed from retail Windows.
  • November 2026: GPT-4 Turbo fallback ends — MAI-Code-1.1-Flash becomes sole default.
  • Watchlist: MAI-Realtime (full-duplex voice) spotted in the MAI Playground but not officially announced — launch likely imminent. MAI-Thinking-1 independent benchmarks still pending broader API access.

Microsoft’s first-party flywheel is now spinning on its own momentum: new models, aggressive pricing, default routing in Copilot, and hard retirement dates for competing infrastructure. The pragmatic stance for engineering teams is to evaluate MAI models for commodity workloads now — the economics and integration depth are becoming a structural advantage that third-party alternatives cannot match on Microsoft platforms.

Follow along at https://x.com/kkaminsk for real-time analysis as these models ship.