Key takeaways
- Microsoft released MAI Code 2 and MAI Code 2 Flash in GitHub Copilot on November 17.
- Microsoft says MAI Code 2 reaches 98% of GPT-6.1 Sol's coding score at about 40% of the cost, and 96% of Claude Opus 5.5 at about 5%.
- They are two of seven models Microsoft AI announced at Ignite, alongside security, image, speech, transcription and routing models.
Microsoft released MAI Code 2 and MAI Code 2 Flash, two coding models built by its own AI lab, on November 17 at its Ignite conference. Microsoft says MAI Code 2 reaches 98% of OpenAI's GPT-6.1 Sol on coding tests at about 40% of the price, and 96% of Anthropic's Claude Opus 5.5 at about 5%. Both models are available now in GitHub Copilot.
They were two of seven models Microsoft AI announced on the same day, covering coding, security, images, speech, medical transcription and routing. The pitch across all of them is the same: close to frontier quality for a fraction of the cost. That matters to any company paying for millions of tokens a day, and it shows Microsoft building models of its own next to the OpenAI ones it resells.
How MAI Code 2 and Code 2 Flash compare
Microsoft's scores are an average over four coding tests: DeepSWE 1.1, Terminal-Bench 2.1, SWE-Atlas and FrontierCode 1.1. All models ran at maximum reasoning effort, through the GitHub Copilot command-line tool, and Microsoft says the FrontierCode scores and costs come from Artificial Analysis.
| Model | Microsoft's claim | Where to get it |
|---|---|---|
| MAI Code 2 | 98% of GPT-6.1 Sol's score at about 40% of the cost; 96% of Claude Opus 5.5 at about 5% | GitHub Copilot |
| MAI Code 2 Flash | Comparable to GPT-6 Luna at 65% of the price, and to Claude Haiku 5.5 at 20% | GitHub Copilot; limited preview in Microsoft Foundry |
Microsoft says MAI Code 2 is meant for everyday enterprise coding: production code, debugging, changes across many files and front-end work. In a blind test of front-end work with 258 reviewers, they preferred MAI Code 2 over GPT-6 Sol 67% of the time. Code 2 Flash is for high-volume jobs like explaining code, small edits and routine fixes.
Microsoft did not publish a per-token price on the announcement page. It says both models are at a 50% promotional price for now, and that they are coming to OpenRouter, Vercel, Hermes, OpenCode and Kilo.
Timing works against the comparison
Every benchmark Microsoft chose sets MAI Code 2 against GPT-6.1 Sol. OpenAI replaced that model the same day with GPT-6.2 Sol, which keeps the same $2 per million input tokens. The 98% figure says nothing yet about the new Sol, and no independent lab had tested MAI Code 2 at publication.
The other five models
- MAI Cyber 2 Flash, for security work. Microsoft says it matches GPT-6 Astra on the CyberGym end-to-end test at under 2% of the cost per task.
- MAI Image 2.7, for making and editing images, in Microsoft Foundry. Microsoft says it ranks first for image editing on Artificial Analysis. A cheaper Flash version costs about $10 per 1,000 images.
- MAI Realtime 1, Microsoft's first live speech-to-speech model, from $0.01 per audio minute.
- Dragon Copilot Transcribe, built on MAI-Transcribe-2, for doctor and patient conversations in Microsoft's Dragon Copilot.
- Decision 1.1, which routes each request to the cheapest model that can handle it. Microsoft says 4,000 customers use it.
Microsoft also announced MAI Licensed Weights, which lets a company take an MAI model, train it further on its own data and run it inside its own Azure account. It named AT&T as an early customer.
What to watch
Microsoft says MAI Code 2 is in GitHub Copilot now, so developers can test the claim on their own code. The open question is price: Microsoft hasn't said what the models cost once the 50% promotion ends, or when that is.


