Two launches in two days. MAI-Code-1.1-Flash shipped August 11 with a 75% cost cut over its predecessor and new image input support. MAI-Image-2.6 landed August 10 at No. 2 on the Arena text-to-image leaderboard, ahead of Google, Meta, and xAI. Both are first-party Microsoft models — trained in-house, routed through Copilot, available across Azure Foundry and GitHub Copilot tiers.

The signal is clear: the MAI family is iterating fast enough to displace third-party models on cost-sensitive workloads while improving quality.

MAI-Code-1.1-Flash: Faster, Cheaper, Multimodal

The headline numbers are stark. Compared to MAI-Code-1-Flash (which shipped at Build 2026 in June), version 1.1 delivers:

  • 75% cost reduction: $0.20/1M input, $0.02/1M cached, $1.20/1M output — down from $0.75/$0.075/$4.50
  • 25% faster token streaming and 25% fewer tokens per task
  • +22% on Terminal-Bench 2.1 (GitHub Copilot CLI): 51.7% → 62.9%
  • +15% on .NET tasks, with +4% code survival and +9% return visits
  • 72.6% on SWE-Bench Verified at 8.6K tokens average per solved task
  • Image input — a first for the MAI-Code line, enabling screenshot-to-code workflows

The architecture remains a sparse MoE at 138B total / ~5B active with 256K context and 128K max output, trained on 2M synthetic tasks and 150K RL environments. It’s now the production model across all GitHub Copilot tiers — Free through Enterprise (Business/Enterprise require admin opt-in). MAI-Code-1-Flash deprecates September 10; GPT-4 Turbo fallback runs through November. If you call MAI-Code via Azure Foundry directly, update your endpoints now.

MAI-Image-2.6: Chasing OpenAI on the Arena

MAI-Image-2.6 launched August 10 and immediately took No. 2 on the text-to-image Arena leaderboard — behind OpenAI but ahead of every other competitor. The gains over MAI-Image-2.5 are broad: +79 Elo overall, +91 Elo on text rendering specifically, with stronger portraits, 3D imagery, and photorealistic commercial outputs.

Available in Arena now; MAI Playground, Foundry, and other surfaces roll out in coming weeks. Microsoft has maintained API compatibility across the 2.x line, so the upgrade path from MAI-Image-2.5 ($47/1M) or Flash ($19.50/1M) is straightforward. The practical implication: Microsoft’s internal data shows an 84% GPU cost reduction versus GPT-Image-2 in production. For image generation at scale, that’s a line item that matters.

MAI-Thinking-1: Public Preview, Independent Benchmarks Still Pending

The flagship reasoning model (~1T total / ~35B active sparse MoE, 256K context) entered public preview on Azure Foundry August 12. Microsoft-reported benchmarks remain impressive: 97.0% on AIME 2025, 94.5% on AIME 2026, 52.8% on SWE-Bench Pro, and a blind preference win over Claude Sonnet 4.6 across 1,276 tasks via Surge.

But independent benchmarks are still scarce. Public preview on Foundry means API access is available for your own evaluation — run your own benchmarks before committing.

Phi-4: The Edge Story Matures

The Phi-4 family now spans 10+ MIT-licensed variants from 3.8B to 15B. The key developments this week:

Azure AI Studio integration (July 25) brought Phi-4 to the cloud with 128K context and Azure ML fine-tuning — practical for hybrid deployments where the same model runs on-device and in the cloud.

Phi-4-mini-flash-reasoning uses a SambaY architecture (Mamba state space + sliding window attention + Gated Memory Units) delivering 10× throughput and 2-3× latency improvement over Phi-4-mini-reasoning. If your edge AI roadmap hasn’t accounted for non-Transformer architectures, this warrants attention.

Phi-4-reasoning-vision-15B (March 2026) is now recommended for new multimodal reasoning workloads — 88.2% on ScreenSpot v2 for GUI grounding, dual reasoning modes, effective replacement for Florence-2 where reasoning is required.

Florence-2: 28 Days Until Hard Deadline

Legacy Azure Vision APIs (v1.0–3.1) retire September 13, 2026 — 28 days from today. Organizations must migrate to Florence-2-based Image Analysis 4.0 SDK. This is a hard cutoff, not a soft deprecation.

No Florence-3 has been announced. The strategy: pure vision stays on Florence-2, reasoning-plus-vision moves to Phi-4-reasoning-vision-15B. Both are MIT-licensed; LoRA fine-tuning yields >98% precision on specialized datasets.

Aion 1.0: The Phi Silica Sunset

Aion 1.0 Instruct is in developer preview (Edge Canary/Dev 150.0.4070+) and will replace Phi Silica as Windows’ on-device model. The transition timeline:

  • October 1: Standalone Aion Instruct testing package
  • October 23: Windows Insider rollout (Phi Silica remains, model selection via CFR)
  • November 24: GA — Aion Instruct becomes production, Phi Silica removed from retail devices

Aion 1.0 Plan (14B, 32K context) is the more ambitious release — an on-device agent runtime with tool-calling and sub-agent orchestration via the Windows Agent Framework, requiring ≥40 TOPS NPU or qualifying discrete GPU. Critically, Aion Instruct doesn’t require LAF tokens unlike Phi Silica — removing a real developer friction point. Phi Silica got a final update in PowerToys v0.101 (August 10), but if you’re building with it today, start your Aion migration plan now.

The Copilot Routing Reality Check

CNBC reported August 5 that Microsoft internally defaulted GitHub Copilot to OpenAI GPT-5.6 Sol for staff use while offering MAI models as options. This isn’t a contradiction — it’s the strategy working as designed. Copilot’s task categorization layer routes commodity tasks to MAI models for cost efficiency; complex reasoning and creative generation go to frontier models. The nuance: first-party models win on high-volume cost-sensitive tasks, frontier models win on complex reasoning. Your architecture should support both — Copilot’s routing layer already does this automatically.

Action Items

  1. Update MAI-Code deployments: If calling MAI-Code via Azure Foundry, migrate to 1.1-Flash before September 10 deprecation of v1.0.
  2. Evaluate MAI-Image-2.6: Image generation at scale — benchmark against your current provider. The cost story is compelling.
  3. Run your own MAI-Thinking-1 benchmarks: Public preview is live on Foundry. Don’t rely on vendor-reported numbers.
  4. Florence-2 migration: September 13 deadline. 28 days. Move now.
  5. Aion migration planning: If building with Phi Silica, the October 1 testing package is four weeks away. Prepare LoRA training data.
  6. Review Copilot routing assumptions: Your Copilot dependency plan should reflect the MAI/third-party split, not assume OpenAI for everything.

Microsoft’s first-party stack spans custom silicon (Maia 200), cloud models (MAI), open-weight edge models (Phi-4), on-device models (Aion), vision foundation (Florence-2), and orchestration (Copilot). Two major model launches in one week — this isn’t a research project anymore. It’s a production platform.

This article is part of the Microsoft AI Weekly series. Follow along at x.com/kkaminsk for ongoing analysis.