Tools & releasesImpact: High
Oleksandr Kuzmenko, EditorPublished 6 min readAI-assisted · editor-reviewed
TL;DR
- MAI Code 2 matches 98% of GPT-6.1 Sol coding benchmark quality at 40% of the price.
- Internal prompt optimization cuts Word token consumption by 42% and Teams by 28%.
- Both models are available in GitHub Copilot and coming to OpenRouter and Vercel.
Key facts
- DeepSWE/SWE-Atlas vs GPT 6.1 Sol
- 98% performance at 40% cost (self-reported)
- DeepSWE/SWE-Atlas vs Claude Opus 5.5
- 96% performance at 5% cost (self-reported)
- Frontend Blind Preference vs GPT-6 Sol
- 64% preference across 258 annotators (self-reported)
- Word Token Efficiency Gain
- ~42% fewer input tokens vs GPT-6.1 Sol
Purpose-Built Architecture for Everyday Engineering. Microsoft has developed MAI Code 2 and MAI Code 2 Flash specifically for routine software engineering workloads. Rather than training models for edge-case reasoning or fundamental scientific research, Microsoft focused on multi-file implementations, repo-wide contextual awareness, deep debugging, and frontend development. In human blind evaluations against GPT-6 Sol covering dashboards, 3D scenes, and SVG components, 64% of 258 annotators preferred MAI Code 2.
Benchmark Performance and Cost Reductions. Averaged across DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas, and FrontierCode v1.1 running via the GitHub Copilot CLI without subagents, MAI Code 2 reached 98% of the performance of GPT-6.1 Sol at 40% of the cost. Compared to Claude Opus 5.5, it retained 96% of the output quality at 5% of the price based on promotional pricing. For high-volume lightweight workflows, MAI Code 2 Flash matches GPT-6 Luna and Claude Haiku 5.5 at 65% and 20% of their operational costs respectively.
Platform Integration and Model Availability. The lighter MAI Work Flash derivative is already serving Microsoft 365 and Fabric IQ. In production Word workloads, it consumes ~42% fewer input tokens per request than GPT-6.1 Sol, while in Teams it follows system instructions using up to 28% fewer tokens. Both MAI Code 2 and MAI Code 2 Flash are immediately accessible inside GitHub Copilot, with preview access in Microsoft Foundry and planned deployments to OpenRouter, Vercel, and Hermes.
Try it in 2 minutes
json
{
"github.copilot.advanced": {
"model": "mai-code-2"
}
}When to use
- Production enterprise code maintenance, multi-file refactoring, and bug resolution.
- Frontend interface generation, dashboard visualization, and SVG creation in GitHub Copilot.
When NOT to use
- Long-horizon autonomous agentic workflows with deeply nested subagent execution.
- Fundamental theoretical research and open-ended scientific problem solving.
What to do today
- Test MAI Code 2 in GitHub Copilot CLI for frontend and visual UI generation tasks.
- Benchmark MAI Code 2 Flash against Claude Haiku for automated inline code explanation pipelines.
- Monitor OpenRouter and Vercel model catalogs for external API availability.
Step names are taken from What to do today.
Step names are taken from What to do today. Test MAI Code 2 in GitHub Copilot CLI for frontend and visual UI generation tasks.. Benchmark MAI Code 2 Flash against Claude Haiku for automated inline code explanation pipelines.. Monitor OpenRouter and Vercel model catalogs for external API availability.
- Test MAI Code 2 in GitHub Copilot CLI for frontend and visual UI generation tasks.
- Benchmark MAI Code 2 Flash against Claude Haiku for automated inline code explanation pipelines.
- Monitor OpenRouter and Vercel model catalogs for external API availability.
What the community says
They are better than the rest of the available models currently, but only by a few months. I think only google is catching up, and others plateaued.
Also “mai” that is “never” in italian
