Microsoft’s MAI-Code-1.1-Flash Brings Vision Coding to Copilot at 73% Lower Cost

Microsoft has added MAI-Code-1.1-Flash to GitHub Copilot, with general availability beginning August 11, 2026. The vision-capable model accepts text and images, opening up workflows such as turning screenshots into web interfaces. LLM Stats reports that it is designed for lower-cost coding assistance rather than frontier-scale reasoning.

The model uses a sparse mixture-of-experts architecture with 138 billion total parameters and 5 billion active, plus a 256K-token context window. GitHub’s list pricing is $0.20 per million input tokens, $0.02 for cached input, and $1.20 per million output tokens. That is roughly 73% below the previous MAI-Code-1-Flash rates, while annual Copilot plans assign it a 0.25x premium-request multiplier.

MAI-Code-1.1-Flash is available through automatic selection for Copilot Free and Student users. Pro, Pro+, Max, Business and Enterprise customers can select it manually or use auto-selection, although Business and Enterprise administrators must enable the model’s policy first. It is listed across Copilot CLI, cloud agent, VS Code, Visual Studio, GitHub Chat and other supported clients.

Microsoft’s model-card results are vendor-reported and used the Copilot evaluation setup. MAI-Code-1.1-Flash scored 72.6% on SWE-Bench Verified, compared with 71.6% for its predecessor, and 62.9% on Terminal Bench 2.1, up from 51.7%. Microsoft also says responses stream 25% faster and use 25% fewer tokens per task.

For AI tool builders, the combination of image input, long context and low usage costs could make lightweight coding agents and screenshot-driven development more practical. The older MAI-Code-1-Flash is scheduled to retire on September 10, 2026, so existing integrations may need updating.

Source: LLM Stats