Microsoft equips Copilot with MAI models designed to reduce costs
Microsoft has unveiled seven MAI models for code, cybersecurity, image, voice, and routing. The goal is clear: to offer performance close to cutting-edge models but at a lower cost per task for businesses.
On November 17, 2026, at Ignite, Microsoft introduced a family of seven MAI models aimed at companies seeking near top-tier results without inference costs at the same level. The set covers code, cybersecurity, image, voice, transcription, and query routing, with a focus on continuous enterprise use. The message is clear: the group wants to win on performance-price ratio, where IT managers look at cost lines before slogans.
Le code, premier terrain de bataille
MAI Code 2 and MAI Code 2 Flash kick off the development side. Microsoft claims Code 2 reaches 98% of GPT-6.1 Sol’s level for 40% of the price, and 96% of Claude Opus 5.5 for 5% of the price, based on the average of four benchmarks: DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas, and FrontierCode v1.1. The Flash version targets quick fixes, targeted edits, and high-volume tasks, with a cost-per-task logic rather than raw record.
On these same tests, Microsoft reports an average score of 63.2% for Code 2, versus 64.1% for GPT-6.1 Sol and 65.3% for Claude Opus 5.5. The announced average cost drops to 0.309 dollars per task for Code 2, compared to 0.761 dollars for GPT-6.1 Sol and 6.306 dollars for Claude Opus 5.5, with prices calculated on a promotional 50% basis. For Code 2 Flash, the group reports a 61.0% pass rate and 0.052 dollars per task, versus 55.1% and 0.080 dollars for GPT-6 Luna, as well as 53.9% and 0.319 dollars for Claude Haiku 5.5.
Microsoft specifies, however, that neither Code 2 nor Code 2 Flash has yet been optimized for fundamental scientific research or very long agentic tasks. The scope is therefore clear: enterprise code, daily use, cost-effective. And, above all, code that does not blow the budget at the end of the month.
Une distribution pensée pour les outils déjà utilisés
Both models are available in GitHub Copilot. Code 2 Flash is also available in limited preview in Microsoft Foundry, with deployments announced on other platforms such as OpenRouter, Vercel, Hermes, OpenCode, and Kilo. GitHub also specifies that Code 2 is supported in Copilot Pro, Pro+, Max, Business, and Enterprise, while Code 2 Flash extends to Free and Student offers on VS Code, Visual Studio, CLI, JetBrains, Xcode, and Eclipse.
This distribution is as important as the benchmarks. Microsoft is not just selling a model; it is pushing a new usage hierarchy within its existing products. In practice, the company seeks to embed MAI into developers’ and IT teams’ already established workflows without forcing them to change application layers.
Cybersécurité, image et voix complètent le tableau
MAI Cyber 2 Flash targets security with a similar approach: fast responses, reduced cost, and a very large context. Microsoft announces 92.3% on CyberGym-E2E, versus 91.5% for GPT-6 Astra, at less than 2% of the cost per task. The model has a context window of one million tokens and is charged 0.10 dollars input and 0.50 dollars output per million tokens. It is set to arrive exclusively in Microsoft security products, replacing the current model in Codename MDASH for eligible customers starting mid-December.
For image, MAI-Image-2.7 ranks first in the AA-Image-Editing v2.1 ranking according to Artificial Analysis. Microsoft also says the Flash version costs about 10 dollars for 1,000 images, versus about 35 dollars for the full model, and is available in public preview via Foundry and several partners, including fal, Genspark, OpenRouter, and Vercel. The Flash version also powers a new Canvas space in MAI Labs.
On the voice side, MAI Realtime 1 becomes Microsoft’s first live speech-to-speech model, with a starting price announced at 0.01 dollars per audio minute and open registration via MAI Labs. Transcription follows the same industrial logic: Dragon Copilot Transcribe, built on MAI Transcribe 2, already processes 10 million patient consultations per month. Here, AI is not in the headlines for fun; it mainly handles volume.
Le routage et le poids stratégique d’Azure
Microsoft-Decision-1.1, the routing model in the series, is now generally available. Its function is simple to describe and useful at large scale: it sends a request to a more powerful model only when necessary. Artificial Analysis assigns it a score of 68 on its Decision Index, versus 62 for the previous version. Microsoft adds that the model remains priced at 0.042 dollars per million input tokens.
The group also launches MAI Licensed Weights, an offer allowing a company to customize a cutting-edge MAI model with its own data and run it in its Azure environment. At the same time, Microsoft indicates that its Maia 200 accelerator is now in production and powers its largest models. This dual software and hardware move clearly shows the intent: better control of the entire chain, from silicon to end use.
By the numbers
- 7 models — new MAI family announced at Ignite 2026.
- 1 million tokens — context window of MAI Cyber 2 Flash.
- 10 million — patient consultations processed monthly by Dragon Copilot Transcribe.
- 0.01 dollars — starting price per audio minute of MAI Realtime 1.
- 68 — Microsoft-Decision-1.1 score on Artificial Analysis’s Decision Index.