Skip to content
, ,

MAI Code 2 in GitHub Copilot: Plans, Price, Benchmarks and How to Use It

9–13 minutes
Developer in a plant-filled workspace holding a tablet showing the GitHub Copilot logo, hook Microsoft's own coding model

Microsoft now has its own coding models inside GitHub Copilot. At Microsoft Ignite on 17 November 2026, Microsoft AI launched MAI Code 2 and the faster MAI Code 2 Flash, and both are already in the Copilot model picker.

The pitch is price. Microsoft says MAI Code 2 gets close to GPT-6.1 Sol and Claude Opus 5.5 on coding tests for a fraction of the cost, and the Flash model is available even on the Copilot Free plan.

This guide covers who gets each model, how to switch to it in VS Code, the CLI and github.com, the real token prices (including the promo that ends in May 2027), the benchmark numbers and their caveats, and when you should still pick a bigger model.

TL;DR

  • What: MAI Code 2 (main coding model) and MAI Code 2 Flash (fast, cheap model for routine edits), made by Microsoft AI, launched 17 November 2026.
  • Where: GitHub Copilot only for MAI Code 2. Flash is also in a limited preview on Microsoft Foundry.
  • Plans: MAI Code 2 on Copilot Pro, Pro+, Max, Business, Enterprise. Flash also on Free and Student.
  • Promo price per 1M tokens (until 17 May 2027): MAI Code 2 $0.50 in / $2.50 out; Flash $0.05 / $0.25. Both double after that.
  • Benchmarks (Microsoft-reported): 63.2% average pass rate vs 64.1% for GPT-6.1 Sol and 65.3% for Claude Opus 5.5.
  • Catch: not tuned for very long agent sessions, optimised for English code, and it uses more tokens per task than Sol.

What Are MAI Code 2 and MAI Code 2 Flash?

MAI stands for Microsoft AI, the in-house model team led by Mustafa Suleyman. Its earlier coding models, MAI Code 1 Flash and 1.1 Flash, were small models for quick Copilot tasks. MAI Code 2 is the first one Microsoft pitches against frontier coding models.

MAI Code 2 MAI Code 2 Flash
Made for Production code, deep debugging, multi-file changes, repo-wide reasoning, UI work Code explanations, targeted edits, routine fixes, fast iteration
Launched 17 Nov 2026 (Ignite) 17 Nov 2026 (Ignite)
Where GitHub Copilot GitHub Copilot + Microsoft Foundry (limited preview)
Microsoft compares it with GPT-6.1 Sol, Claude Opus 5.5 GPT-6 Luna, Claude Haiku 5.5
Open weights No No
Context window Not published by Microsoft at launch

Microsoft says both models will also come to OpenRouter, Vercel, Hermes, OpenCode and Kilo later, with no dates yet. Large companies can also customise an MAI model with their own data and run it in their Azure account through the “MAI Licensed Weights” offer.

Note: Flash already powers other Microsoft products. A version called MAI Work Flash, built from Code 2 Flash, now runs some Microsoft 365 features in Word, Excel, PowerPoint and Teams, according to Microsoft.

Which Copilot Plans and Apps Get Them

Copilot plan MAI Code 2 MAI Code 2 Flash
Free No Yes
Student No Yes
Pro Yes Yes
Pro+ Yes Yes
Max Yes Yes
Business Yes (if admin policy allows) Yes (if admin policy allows)
Enterprise Yes (if admin policy allows) Yes (if admin policy allows)

Microsoft and GitHub say access is for “eligible” subscribers on usage-based billing, and reports mention annual plans. If you are on an older request-based plan, check your billing page.

Apps that show the models

  • VS Code and Visual Studio
  • JetBrains IDEs, Xcode and Eclipse
  • GitHub Copilot CLI
  • Copilot cloud agent and the GitHub Copilot app
  • github.com chat
  • GitHub Mobile on iOS and Android

How to Use MAI Code 2 in Copilot, Step by Step

In VS Code (same idea in JetBrains and Visual Studio)

  1. Update VS Code and the GitHub Copilot Chat extension. GitHub’s advice if a model is missing is simply to update the client.
  2. Open Copilot Chat from the sidebar or with the chat shortcut.
  3. Click the model picker under the chat box and choose MAI Code 2 for real feature work, or MAI Code 2 Flash for quick questions and small edits.
  4. Use Agent mode for multi-file changes. Keep each task scoped (one feature or one bug), since Microsoft says the model isn’t tuned for very long agent runs.
  5. Review the diff before you accept it, and run your tests.

In the Copilot CLI and on github.com

  • CLI: open the model selector inside the Copilot CLI session and pick MAI Code 2. Microsoft ran its own benchmarks through the Copilot CLI, so this is the setup its numbers describe.
  • github.com: open Copilot chat on any repository page and choose the model from the dropdown.
  • Cloud agent: when you assign an issue to Copilot, choose the model in the agent settings if your plan shows it.

For Business and Enterprise admins

  1. Go to your organisation or enterprise Copilot settings and open the model policy.
  2. Check that MAI Code 2 and Flash are enabled. With default model enablement on, new models turn on automatically unless you disabled the global default or the model itself.
  3. Set a budget for AI credits if you want a spending ceiling (see the next section).

Tip: using Copilot in Xcode? Our GitHub Copilot in Xcode setup guide shows where the model picker lives there.

Pricing, AI Credits and the May 2027 Price Rise

Copilot now bills model use in AI credits, where 1 credit = $0.01. Input, cached and output tokens are converted into credits. Code completions are not billed in credits, and credits reset monthly without rolling over.

Per 1M tokens MAI Code 2 (promo) MAI Code 2 (from 17 May 2027) Flash (promo) Flash (from 17 May 2027)
Input $0.50 $1.00 $0.05 $0.10
Cached input $0.025 $0.05 $0.005 $0.01
Output $2.50 $5.00 $0.25 $0.50

Microsoft’s own launch post labels these as a 50% promotional price. Flash’s Foundry price has not been published yet.

What a task actually costs

  • Microsoft’s average cost per benchmark task: $0.309 for MAI Code 2 at promo prices vs $0.761 for GPT-6.1 Sol.
  • At the full price from May 2027, eesel estimates about $0.618 per task, roughly 81% of Sol’s cost.
  • The gap is smaller than the token price suggests because MAI Code 2 used about 1.6 times as many tokens per task as Sol in those runs.

How to cap your spend

  1. Open your GitHub billing settings and find the Bundled AI credits budget.
  2. Create or edit the budget and set the amount you are happy to pay (set it to $0 to use only included credits).
  3. Turn on “Stop usage when budget limit is reached”.

Warning: Business and Enterprise allowances are pooled across the billing entity, so one heavy user can eat into everyone’s credits. Set budgets before rolling the model out widely.

Benchmarks and What They Really Show

All the numbers below are Microsoft-reported. Microsoft averaged four coding benchmarks: DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1. Every model ran at maximum reasoning effort in the GitHub Copilot CLI, without subagents.

Measure MAI Code 2 GPT-6.1 Sol Claude Opus 5.5
Average pass rate (4 benchmarks) 63.2% 64.1% 65.3%
SWE-Atlas 43.1% – 57.5%
Microsoft’s headline – 98% of Sol’s score at 40% of the cost 96% of Opus’s score at 5% of the cost
Cost per task (promo) $0.309 $0.761 –

Two more Microsoft claims:

  • In blind front-end tests with 258 developers, MAI Code 2’s results were preferred over GPT-6 Sol 64% of the time.
  • MAI Code 2 Flash is “comparable” to GPT-6 Luna at 65% of its cost and Claude Haiku 5.5 at 20% of its cost, on three everyday coding benchmarks. Microsoft gave no single score.

The caveats

  • Mostly in-house tests: eesel notes three of the four benchmarks were run by Microsoft; only FrontierCode came from an outside group.
  • Already-old comparison: GPT-6.1 Sol was replaced by GPT-6.2 Sol on the same day, at the same price.
  • Weak spot: the 14.4-point gap to Opus on SWE-Atlas suggests it is weaker at deep repository questions.
  • No independent leaderboard scores yet from groups like Artificial Analysis or Vals AI at the time of writing.

MAI Code 2 vs GPT-6.2 Sol vs Claude Opus 5.5 vs Haiku 5.5

MAI Code 2 MAI Code 2 Flash GPT-6.2 Sol Claude Opus 5.5
Maker Microsoft AI Microsoft AI OpenAI Anthropic
Price per 1M (in / out) $0.50 / $2.50 promo $0.05 / $0.25 promo $2 / $10 Much higher (Microsoft: MAI Code 2 costs ~5% as much per task)
Best for Scoped multi-file work on a budget Explanations, small fixes Everyday coding and agents Long, hard repo tasks
Outside Copilot No (coming to OpenRouter etc.) Foundry preview OpenAI API, ChatGPT, Codex Claude API, Claude Code

GPT-6.2 Sol price from OpenAI’s model page (see our OpenAI API pricing guide); MAI figures from Microsoft and GitHub. If you code mostly in a terminal agent rather than Copilot, our Codex vs Claude Code comparison and Cursor vs Windsurf vs Claude Code vs Cline cover the alternatives.

Problems and Fixes

Problem Likely cause Fix
MAI Code 2 not in the model picker Old client, Free/Student plan, or admin policy Update the IDE and Copilot extension; check your plan; ask your admin to enable it
Only Flash shows up You are on Copilot Free or Student Upgrade to Pro or higher for MAI Code 2
Usage stops mid-month AI credits budget reached Raise the Bundled AI credits budget or wait for the monthly reset
Agent drifts on long tasks Microsoft’s model card says long sessions can drift Split the job into smaller tasks, or switch to a larger model for long runs
Weaker results in non-English code comments or docs Optimised for English code Write prompts in English, or use a general-purpose model for other languages
Can’t call it from your own app MAI Code 2 is Copilot-only for now Use Flash in the Foundry preview if you have access, or wait for OpenRouter/Vercel

Limits, Privacy and What Microsoft Hasn’t Said

  • Not for long autonomous work: Microsoft says the coding models are “not yet tuned for fundamental science research or very long-horizon agentic work”.
  • Closed model: no weights for download and no local option. Only enterprise customers can get licensed weights to run in Azure.
  • Your code: it runs inside Copilot, so your existing Copilot data and privacy settings apply. Business and Enterprise policies control which models people can use.
  • Unknowns: context window, knowledge cutoff, model size, and independent benchmark results were not published at launch.

My Take

This is the first time Microsoft has put its own model up against OpenAI and Anthropic inside GitHub’s flagship product, and that matters more than any single benchmark. Copilot no longer depends only on partners.

For everyday Copilot users, the practical win is cost. During the promo, MAI Code 2 is cheap enough to be a sensible default for scoped feature work, and Flash on the Free plan is a real upgrade for students and hobbyists.

I would not make it the default for long, hard refactors yet. The numbers are mostly Microsoft’s own, the SWE-Atlas gap is large, and the token-hungry behaviour cuts into the savings.

Try it on your next few small tasks, keep a bigger model for the tough ones, and look again before the price doubles on 17 May 2027.

FAQ

Is MAI Code 2 free?

MAI Code 2 needs Copilot Pro or higher. MAI Code 2 Flash is available on the Copilot Free and Student plans, within those plans’ usage limits.

Who made MAI Code 2?

Microsoft AI, Microsoft’s in-house model team. It is not an OpenAI model, although it runs alongside OpenAI and Anthropic models in Copilot.

How much does MAI Code 2 cost?

$0.50 per million input tokens and $2.50 per million output tokens at the promotional rate until 17 May 2027, then $1.00 and $5.00. In Copilot, this is charged as AI credits at $0.01 each.

Can I use MAI Code 2 outside GitHub Copilot?

Not yet. Only Flash is in a limited Microsoft Foundry preview. Microsoft says OpenRouter, Vercel, OpenCode, Kilo and Hermes will get the models later.

Is MAI Code 2 better than Claude Opus 5.5?

No, not on Microsoft’s own numbers. It scores 63.2% on average vs 65.3% for Opus 5.5, but Microsoft says it costs about 5% as much per task.

Can I run MAI Code 2 locally?

No. The weights are not public. Enterprise customers can license weights to run in their own Azure environment.

What is the context window of MAI Code 2?

Microsoft did not publish a context window at launch. We will add it when the model card lists it.

Sources

  • Microsoft AI, “Introducing MAI Code 2 and MAI Code 2 Flash” (17 Nov 2026)
  • WindowsForum, “GitHub Copilot Adds MAI Code 2 and Flash: Plans, Pricing and Admin Controls”
  • eesel AI, “MAI Code 2: pricing, benchmarks, and who should use Microsoft’s coding model” and “MAI Code 2 pricing: promo rates, Copilot credits, and cost per task”
  • NeoTeo, “Microsoft’s cheaper AI models: coding claims explained”
  • OpenAI developer docs, gpt-6.2-sol model page (price)
  • No Code MBA AI release tracker, November 2026
Free weekly email

Get new AI guides every week

One short weekly email with our newest local AI, ComfyUI and AI tool guides. No spam, no pop-ups, just the good stuff.

Free. Unsubscribe anytime in one click. We only use your email to send this newsletter. See our privacy policy.

Developer in a plant-filled workspace holding a tablet showing the GitHub Copilot logo, hook Microsoft's own coding model