Microsoft Ships MAI-Code-1.1-Flash, Cutting GitHub Copilot Coding-Model Costs 73% and Adding Vision Support

Overview

Microsoft began rolling out MAI-Code-1.1-Flash, an updated version of its in-house small-tier coding model, in GitHub Copilot on August 11, according to GitHub’s changelog. The company says the new model “adds native vision support for image understanding and delivers improvements across coding quality, instruction following, tool use, and performance” over its predecessor, and lists it at “73% lower list price than MAI-Code-1-Flash,” according to GitHub.

Microsoft AI describes the update as delivering “higher quality code, at 25% greater token efficiency, and at a quarter of the cost” compared with the prior version, according to its own announcement.

What We Know

MAI-Code-1.1-Flash builds on MAI-Code-1-Flash, the coding model Microsoft previously reported launched in June as one of seven proprietary MAI models, where it was described as a model with “5 billion active parameters” built to be “tailor-made for and deeply integrated into GitHub Copilot, VS Code and the Microsoft stack.”

On the benchmark side, Microsoft AI reports “22% improvement on Terminal-Bench 2.1 for GitHub Copilot CLI” and “15% improvement on .NET tasks,” along with a “4% rise in code survival rates” and a “9% increase in return visits,” according to the company’s announcement. The same post states that “in GitHub Copilot tokens stream 25% faster and the model uses 25% fewer tokens to complete a task,” and describes the model as “a lightweight, agentic model built into GitHub Copilot and VS Code,” according to Microsoft AI.

On pricing and billing, annual Copilot subscribers are charged a “0.25× premium request multiplier” for the model, according to GitHub’s changelog. Copilot Free and Student users get the model through automatic selection, while Pro, Pro+, Max, Business, and Enterprise subscribers can select it manually, per the same changelog. The rollout spans Copilot CLI, the Copilot cloud agent, the GitHub Copilot app, Copilot Chat, VS Code, Visual Studio, GitHub Mobile, JetBrains IDEs, Eclipse, and Xcode, according to GitHub. Copilot Enterprise and Copilot Business administrators must separately enable the model through a Copilot settings policy that is “off by default,” GitHub said.

GitHub is simultaneously retiring the predecessor model. In a companion changelog post, the company said “with the launch of MAI-Code-1.1-Flash, we will deprecate MAI-Code-1-Flash,” setting a cutoff of “September 10, 2026” and telling customers to “update your workflows and integrations to use the new model before September 10,” according to GitHub’s deprecation notice. GitHub added that “no action is required to remove the older model once it has been deprecated.”

What We Don’t Know

Neither Microsoft’s announcement nor GitHub’s changelog specifies the model’s parameter count, context window, or the exact technical mechanism behind its vision support. Microsoft AI attributes the gains to optimizing “for real-world use across more than hundreds of thousands of reinforcement-learning environments in GitHub Copilot,” but neither post details the training methodology further, nor do they publish per-token dollar pricing or a comparison against third-party coding models such as Claude, GPT, or Gemini variants also available in Copilot.

Analysis

The update lands as GitHub’s Copilot model picker keeps growing: Moonshot AI’s Kimi K3 went generally available on August 6, as The Machine Herald previously reported, Google’s Gemini 3.7 Flash began rolling out on August 13, according to GitHub’s changelog, and xAI’s Grok 4.6 followed on August 14, according to GitHub. Pairing a steep list-price cut with concrete benchmark gains on its own small-tier model gives Microsoft a cost-efficiency argument to make inside an IDE surface where developers can now switch between competing models with a single click — and where MAI-Code-1-Flash’s own upcoming shutdown leaves existing integrations with a hard September 10 deadline to migrate.