Microsoft launches MAI Code 2 and MAI Code 2 Flash on GitHub Copilot
The third generation of Microsoft's in-house coding models arrives with bold cost claims and a limited Foundry preview for the Flash variant
Microsoft is rolling out MAI Code 2 and MAI Code 2 Flash, the newest members of its homegrown coding model family, on GitHub Copilot. The announcement landed on November 17, 2026.
The Flash variant is reported to have beaten GPT-6.1 in AT&T evaluations while costing 50x less.
What Microsoft actually shipped
MAI Code 2 and MAI Code 2 Flash are both headed for GitHub Copilot. MAI Code 2 Flash is also sitting in a limited-access preview inside Microsoft Foundry.
The headline performance claim comes from testing at AT&T. In those evaluations, MAI Code 2 Flash was reported to outperform GPT-6.1, with the research summary specifying certain GPT-6.1 variants rather than the whole family.
MAI Code 2 Flash is reported to be 50x cheaper than the competition in those tests.
One caveat sits over all of this. Independent benchmarks have not verified these performance claims, and details about wider availability beyond the Foundry preview remain unconfirmed.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
How Microsoft got here
The series started in June 2026 with MAI Code 1 Flash. That model was the first coding model Microsoft built in-house for Copilot.
Two months later, in August 2026, MAI Code 1.1 Flash followed. It added native vision support, meaning the model could work with images and not just text, and it came with pricing roughly 73% lower.
Those earlier releases reportedly delivered higher code accept rates, which measure how often developers actually keep the suggestions the model offers, along with lower token usage.
What this means for developers and the AI coding race
The research frames this as a move away from leaning on third-party architectures such as those from OpenAI and Anthropic. Owning the model means owning the cost structure, and owning the cost structure means more control over what Copilot users ultimately pay.
For enterprises, the AT&T testing is the most interesting detail. It suggests Microsoft is courting large organizations whose coding bills scale with headcount and usage, where a 50x cost difference compounds quickly.
The risks are equally clear. Vendor-run evaluations are a starting point, not a verdict, and claims about beating certain GPT-6.1 variants leave room for interpretation about which variants and which tasks.
The limited preview also means real-world data from a broad user base is not yet available.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.

