DEV Community

News
News

Posted on

Microsoft AI ships 7 in-house models at Ignite 2026 to cut external-model reliance

Infographic cover

๐Ÿ“… November 17, 2026 โ€ข Microsoft AI Ignite / The Decoder

Microsoft AI used Ignite 2026 to ship seven in-house models in one batch โ€” MAI Code 2, MAI Code 2 Flash, MAI Cyber 2 Flash, MAI Image 2.7, MAI Realtime 1, Dragon Copilot Transcribe and Decision 1.1. The framing from The Decoder is direct: Microsoft is replacing external models inside its own products with cheaper in-house alternatives.

The flagship is MAI Code 2. By Microsoft's numbers it delivers 98% of GPT-6.1 Sol performance at 40% of the price and 96% of Claude Opus 5.5 at 5% of the price, averaged across DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1. In Microsoft's blind frontend test with 258 annotators, Code 2's work was preferred over GPT-6 Sol in 67% of cases. GitHub Copilot promo pricing runs through May 17, 2027 โ€” Code 2 at $0.50/$2.50 per 1M input/output tokens (then $1/$5), Code 2 Flash at $0.05/$0.25 (then $0.10/$0.50), with Flash pitched against GPT-6 Luna at 65% of its price and Claude Haiku 5.5 at 20%.

MAI Cyber 2 Flash scores 92.3% on CyberGym-E2E against 91.5% for GPT-6 Astra, which needed twice the reasoning budget, at a per-task cost under 2% of Astra's โ€” per Microsoft. It carries a 1M-token context but ships only inside Microsoft security products, replacing external models in MDASH from mid-December. MAI Image 2.7 ranks #1 in image editing on Artificial Analysis, ahead of GPT Image 2.6, at roughly one-fifth of its price, and #4 in text-to-image. MAI Realtime 1, Microsoft's first live speech-to-speech model, starts at $0.01 per minute, Dragon Copilot Transcribe (on MAI Transcribe 2) handles ~10M patient consultations a month, and Decision 1.1 routes between models at $0.042 per 1M input tokens.

Inside Microsoft, MAI Work Flash, built out from Code 2 Flash, already serves Excel, PowerPoint, Word, Teams and FabricIQ, cutting input tokens per call by ~42% in Word versus GPT-6.1 Sol in production and up to 28% in Teams, according to Microsoft. The MAI Licensed Weights program lets enterprises customize MAI models with their own data and run them in their own Azure โ€” "No one wants to pour their IQ into someone else's model," as the company puts it. The second-generation Maia 200 chip is in production running Microsoft's largest models. Code 2 is live in GitHub Copilot Pro, Pro+, Max, Business and Enterprise, Code 2 Flash also covers Free and Student tiers plus a limited Microsoft Foundry preview, with OpenRouter, Vercel, Hermes, OpenCode and Kilo to follow. Microsoft itself flags the limits: all comparisons are its own, computed at promo prices with a 50% discount baked in, Microsoft has published no independent benchmark results, the models are hosted rather than downloadable weights, and Code 2 is not tuned for fundamental science or long agentic tasks.

Key figures

IMPACT: for automation and agent builders this is primarily a cost-curve move. Always-on agents โ€” code review, monitoring, triage, transcription โ€” get cheaper to run near frontier quality, per Microsoft's numbers. It also reinforces the model-per-task pattern: a Flash-tier model for high-volume calls, a flagship for hard reasoning, routing in the style of Decision 1.1 to orchestrate. And Licensed Weights plus in-house silicon make control over weights and deployment location part of vendor selection.

Sources: microsoft.ai, the-decoder.com, mlq.ai, zdnet.com, github.blog, itwire.com

IMPACT checklist


Continue the conversation: https://t.me/AIAutomationArchitecture ยท Bluesky

AI Automation

Top comments (0)