Microsoft’s in-house AI group used Ignite 2026 on November 17 to release a family of seven models. The goal isn’t to top the frontier leaderboards. Microsoft’s pitch is “a price point that lets you run agents 24-7 doing real engineering work”. The two coding models went straight into GitHub Copilot the same day, so most developers can try them without signing up for anything new.
The two coding models
MAI Code 2 is the general-purpose model. Microsoft aims it at deep debugging, multi-file implementation, refactoring, repo-wide reasoning and UI work. MAI Code 2 Flash is the lightweight one, for code explanations, targeted edits, routine fixes and fast iteration. Both take text and images. AI Ping, citing Microsoft, reports a 1.05M-token context window for each.
Microsoft is direct about what they’re not for. They’re “explicitly not yet tuned for harder problems, like fundamental science research and very long horizon agentic work”. Read the benchmarks with that in mind.
Pricing in Copilot (per 1M tokens)
| Model | Input | Cached input | Output | From May 17, 2027 |
|---|---|---|---|---|
| MAI Code 2 | $0.50 | $0.025 | $2.50 | $1.00 / $0.05 / $5.00 |
| MAI Code 2 Flash | $0.05 | $0.005 | $0.25 | $0.10 / $0.01 / $0.50 |
| GPT-6.2 Sol (for comparison) | $2.00 | $0.10 | $10.00 | — |
| Claude Haiku 5.5 (≤100K) | $0.10 | $0.01 | $0.50 | — |
Copilot bills these in AI credits (1 credit = $0.01). Even at the post-promo rate, Code 2 is half the price of GPT-6.2 Sol. At its promo rate, Flash costs half of Haiku 5.5.
Availability
- MAI Code 2: Copilot Pro, Pro+, Max, Business and Enterprise.
- MAI Code 2 Flash: all of the above plus Free and Student.
- Clients: VS Code, Visual Studio, JetBrains, Xcode, Eclipse, Copilot CLI, Copilot cloud agent, the Copilot app, github.com and mobile. Business/Enterprise admins control access through the model policy. New models are on by default unless an admin has turned that off.
- Outside Copilot, Flash is in limited preview in Microsoft Foundry. Microsoft says OpenRouter, Vercel, OpenCode and Kilo are “soon”.
What the benchmarks say
All results are Microsoft’s, run at max reasoning effort through the Copilot CLI with subagents disabled:
- Code 2: averaged over DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1, it gets 98% of GPT-6.1 Sol’s score at ~40% of the cost and 96% of Claude Opus 5.5’s at ~5%. Note that the cost comparison uses Code 2’s 50% promo price. In a blind frontend evaluation with 258 annotators, reviewers picked Code 2 over GPT-6 Sol 64% of the time.
- Code 2 Flash: on DeepSWE 1.1, Terminal Bench 2.1 and SWE-Atlas, it’s comparable to GPT-6 Luna at 65% of the price and Claude Haiku 5.5 at 20%.
None of this is independently verified yet. “98% of the score” is also an average, so it can hide a much bigger gap on your hardest tasks.
The rest of the Ignite lineup
- MAI Cyber 2 Flash. Microsoft says it matches GPT-6 Astra on CyberGymE2E at under 2% of the cost per task, with output tokens 86% cheaper than Cyber 1.1 Flash. It’s built for continuous defensive scanning.
- Decision 1.1. A routing model that sends requests to bigger models only when they’re needed. Microsoft claims it’s 50x faster and 79x cheaper than GPT-6.1 Sol and has 4,000 customers.
- MAI Image 2.7 (#1 on Artificial Analysis image editing), MAI Realtime 1 speech-to-speech from $0.01 per audio minute, and MAI Transcribe 2 in Dragon Copilot.
- MAI Licensed Weights. Enterprises can take a MAI model, fine-tune it on their own data and run it in their own Azure environment.
- Maia 200, Microsoft’s second-generation accelerator, is now serving its largest models at 1.7x more tokens per watt.
Should you switch?
For high-volume, low-risk work, it’s worth a test this week: explanations, small fixes, test scaffolding, subagent grunt work. Flash at $0.05 / $0.25 is about as cheap as a capable coding model in Copilot gets. Code 2 is a sensible default for everyday multi-file changes. Microsoft itself says to escalate long-horizon or research-heavy tasks to a frontier model. Before you change the default for a whole team, run a few of your own real tickets through both models.