Microsoft releases MAI Code 2 coding model: 98% of GPT-6.1 Sol's quality for 40% of the price
At Ignite 2026, Microsoft showed seven models. The one that matters most for developers is MAI Code 2 in GitHub Copilot. On four coding benchmarks it delivers 98% of GPT-6.1 Sol's quality at 40% of the price. The caveat: that price reflects a 50% promo discount, and at the regular rate it is about 80%. In practice, tasks like editing a function or making sense of someone else's code can go to a cheaper model instead of a flagship.
AI-processed from TG: @ai_machinelearning_big_data (Machinelearning); edited by Hamidun News
At Ignite 2026, Microsoft introduced seven in-house models, and the story here is price, not record-setting intelligence. The most notable release for developers, MAI Code 2, delivers 98% of GPT-6.1 Sol's quality on four coding benchmarks while costing 40% as much. Put simply: if you ask an AI every day to fix a function, write a test, or work out someone else's module, most of those tasks can be done noticeably cheaper with little loss in results. There is an important caveat. Microsoft calculates the price using a 50% promo discount, so at the regular rate the share doubles, to about 80%. Below, we go through what's in the lineup and what to do with it in practice.
Coders: almost flagship quality, lower price
MAI Code 2 is built for everyday enterprise development. It is scored on the average result of four benchmarks: DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode 1.1. On these, the model delivers 98% of GPT-6.1 Sol's quality at 40% of its price, and 96% of Opus 5.5's quality at 5% of its price. At the regular rate, without the discount, those figures are roughly 80% and 10%. In a blind comparison of frontend code, developers chose it over GPT-6 Sol in 68% of cases. The model is available in GitHub Copilot on plans from Pro to Enterprise.
MAI Code 2 Flash is the smaller version for frequent, minor tasks: explaining code, making a targeted edit, fixing a bug. On three benchmarks from the same list, it performs at roughly the level of GPT-6 Luna at 65% of its price, and Haiku 5.5 at 20%, again at the promo price. It is available on all Copilot plans, including the free one, plus a limited preview in Foundry.
The practical takeaway: routine work like editing a function doesn't need an expensive model. Keep the larger version for more serious tasks and point Flash at the high volume of small jobs.
Security, images, voice and healthcare
MAI Cyber 2 Flash protects code and infrastructure. It has a 1M-token context window, four times that of its predecessor, so large codebases fit inside it. On the CyberGym-E2E benchmark it scored 92.3% against 91.5% for GPT-6 Astra, and the cost per task came to less than 2% of the competitor's. Pricing is 10 cents per million input tokens and 50 cents per million output tokens. The model is available only in Microsoft's security products. In MDASH, it will replace the current model starting in mid-December.
MAI Image 2.7 generates and edits images, including from five references. In the Artificial Analysis image-editing leaderboard, the new model took first place and beat GPT Image 2.6 Sunburst, while costing one-fifth as much. The Flash version is built for high-volume workloads and costs 10 dollars per million image output tokens. It is available in MAI Labs, Foundry, OpenRouter and on other partner platforms.
MAI Realtime 1 is Microsoft's first model for live voice conversation: it takes speech in and responds with speech right away. On external benchmarks and on Dynamics 365 tests, it outperforms GPT-Realtime 2.1. Pricing starts at 1 cent per minute of audio, with reasoning billed separately. For now it is a limited preview in MAI Labs, for personal, non-commercial use only. Businesses can submit a request.
Dragon Copilot Transcribe runs on MAI-Transcribe-2 and transcribes all doctor-patient conversations inside Dragon Copilot. It is not sold separately; it comes as part of Dragon Copilot.
Router and open weights
Microsoft-Decision-1.1 handles utility tasks inside applications: it routes requests, classifies, sets priorities and checks results. It is built on Gemma 4 26B A4B and understands images. On the Artificial Analysis Decision 2.1 index, the model scores 68 against 62 for the previous version and runs twice as fast at the same price. General access is open in Foundry, OpenRouter and Vercel. A real-world example: such a model first decides that a request is simple and sends it to a cheap model, while passing a complex one to an expensive model.
For companies that need more than a ready-made API, there is MAI Licensed Weights, the weights of the MAI family. They can be fine-tuned on a company's own data and run in its own Azure environment.
What to keep in mind
Nearly all the figures in the announcement come from Microsoft itself, and the prices in the comparisons reflect the promo discount. Before switching models in production, run them on your own tasks and calculate the cost at the regular rate. Some of the new releases are still limited: Cyber 2 Flash is available only in Microsoft's security products, and Realtime 1 is only a closed preview.
The big picture: Microsoft is building its own model family that covers typical tasks more cheaply than the market leaders. For developers, that's a reason to rethink their setup: Flash for the stream of small jobs, MAI Code 2 for the main workload, and expensive models only where they are truly needed.
- Nov 18, 2026Willison Releases Three External Sandboxes So Agent-Run Code Stays Off Your Laptop
- Nov 18, 2026Nvidia's data-center revenue jumps 110% to $107.6 billion: the AI pause didn't hurt
- Nov 18, 2026Decision models took over 25% of requests in Vercel's router a month after launching
- Nov 18, 2026Vercel releases eve 1.0: agent turn latency drops from 3.04 s to 0.4–0.6 s
Key Questions Answered
What are the key details on Coders: almost flagship quality, lower price?
MAI Code 2 is built for everyday enterprise development. It is scored on the average result of four benchmarks: DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode 1.1. On these, the model delivers 98% of GPT-6.1 Sol's quality at 40% of its price, and 96% of Opus 5.5's quality at 5% of its price. At the regular rate, without the discount, those figures are roughly 80% and 10%. In a blind comparison of frontend code, developers chose it over GPT-6 Sol in 68% of cases.
What are the key details on Security, images, voice and healthcare?
MAI Cyber 2 Flash protects code and infrastructure. It has a 1M-token context window, four times that of its predecessor, so large codebases fit inside it. On the CyberGym-E2E benchmark it scored 92.3% against 91.5% for GPT-6 Astra, and the cost per task came to less than 2% of the competitor's. Pricing is 10 cents per million input tokens and 50 cents per million output tokens. The model is available only in Microsoft's security products. In MDASH, it will replace the current model starting in mid-December.
What are the key details on Router and open weights?
Microsoft-Decision-1.1 handles utility tasks inside applications: it routes requests, classifies, sets priorities and checks results. It is built on Gemma 4 26B A4B and understands images. On the Artificial Analysis Decision 2.1 index, the model scores 68 against 62 for the previous version and runs twice as fast at the same price. General access is open in Foundry, OpenRouter and Vercel.
Build & Ship Production AI Agents Without Boilerplate Code
Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.