Skip to content
AI & Automation

MAI Code 2 explained: what it changes for small teams

Nov 18, 2026

MAI Code 2 is Microsoft AI's general-purpose coding model, and GitHub made it available in GitHub Copilot on November 17, 2026, alongside a lighter model called MAI Code 2 Flash (GitHub's launch changelog). Against the prior Flash model, MAI-Code-1.1-Flash, the change is a higher per-token rate for the large model, a lower rate for the Flash model, and a plan rule that keeps the large model off the Free and Student plans.

The decision this month is not whether to buy a new coding tool. It is which model your Copilot admins allow by default, and whether the heavier model earns its higher rate on your work.

A coding model is the model Copilot sends your code requests to when you pick it in the model picker, and GitHub says both new models are billed under usage-based billing.

TL;DR: Use MAI Code 2 for multi-file and repository-wide work where its higher rate pays off, use Flash for routine edits, and check plan access and admin policy before anyone switches.

Key Takeaways

  • MAI Code 2 is listed for Copilot Pro, Pro+, Max, Business and Enterprise, so users on Free or Student cannot select it.

  • MAI Code 2 Flash is listed for every Copilot plan, including Free and Student, which makes it the lowest-friction model to roll out.

  • Per token, MAI Code 2 is priced above the prior Flash model and MAI Code 2 Flash is priced below it, so the two models pull the bill in opposite directions.

  • Every performance figure for MAI Code 2 comes from Microsoft, and our search found no third-party benchmark of it.

  • Business and Enterprise admins should check the model policy now, because GitHub's default setting enables new models automatically unless an admin turns that off.

  • Treat the promotional rate as temporary, since the post-promotion rates are higher than the current ones.

What launched, and what it replaces

GitHub describes MAI Code 2 as built for "deep debugging, multi-file implementation, refactoring, and repository-wide reasoning," and MAI Code 2 Flash as "optimized for everyday development workflows," such as code explanations, targeted edits, routine fixes and rapid iteration (GitHub's launch changelog). Both are described as designed for agentic coding workflows. GitHub points to Microsoft's announcement for vision details, but the page we opened does not state which inputs are supported.

One naming point needs flagging. GitHub's billing docs now list MAI-Code-1.1-Flash and do not list MAI-Code-1-Flash (GitHub's billing docs), and Microsoft's 1.1 page refers to the June launch as "1.0" (Microsoft AI's 1.1 page). We found no page that says the June model was discontinued or merged. The honest reading is that the Flash line is now billed under the 1.1 name, not that a model was retired. Confirm which name your model picker shows before you write any policy.

Plan access differs by model. GitHub lists MAI Code 2 for Copilot Pro, Pro+, Max, Business and Enterprise, and MAI Code 2 Flash for Copilot Free, Student, Pro, Pro+, Max, Business and Enterprise. The changelog adds that both are available to eligible subscribers on annual plans. That line needs checking, because the plans page we opened does not show annual prices. For the earlier Flash model, GitHub's June 26, 2026 changelog says it became generally available for Copilot Business and Enterprise, and that an administrator "must enable the MAI-Code-1-Flash policy in Copilot settings before users can access it" (GitHub changelog, June 26, 2026).

For the admin setting on the new models, GitHub says Business and Enterprise administrators manage access through the model policy in Copilot settings, and that under default model enablement, new models are enabled automatically unless an administrator has turned off the global default or disables the model directly (GitHub's launch changelog). That is a different default from the earlier Flash model's opt-in policy, which is why the admin check belongs at the top of your list.

For a parallel launch breakdown on another model, see our Composer 2.5 explainer.

How we evaluated these tools

We scored the two new models and the prior Flash model on six criteria that a small B2B team decides on, using the weights below. The weights are our judgment, not vendor data. Plan access and cost carry the most weight, because they decide whether a model is usable this month at all.

CriterionWeightWhy it matters for this reader
Plan access and admin control25%A model your plan or admin policy blocks is not an option, whatever its benchmark
Per-token price and plan cost25%Sets the monthly run rate once engineers use a model daily
Fit to the task mix20%Multi-file work and routine edits reward different models
Quality of evidence15%Vendor-only claims score lower than independent checks
Switching effort10%Defaults, policy changes and team retraining take time
Disclosed limits and unknowns5%Missing context and vision details make planning harder

Evidence quality carries 15% because almost every performance figure for these models comes from the vendor. A model can score well on the criteria that matter to a buyer and still need a verification step before it becomes a default.

Side-by-side on the decisions that matter

The table below sets the three models against the decisions a buyer makes. Figures that are not on the vendor pages we opened are marked as not stated.

Decision pointMAI Code 2MAI Code 2 FlashMAI-Code-1.1-Flash (prior)Source
Intended jobDeep debugging, multi-file work, refactoringExplanations, targeted edits, routine fixesSmall, efficient coding modelGitHub launch, Microsoft AI 1.1
Plans with accessPro, Pro+, Max, Business, EnterpriseFree, Student, Pro, Pro+, Max, Business, EnterpriseFree, Student, Pro, Pro+, Max; Business and Enterprise added June 26, 2026GitHub launch, GitHub June 18, GitHub June 26
Admin defaultNew model enabled unless an admin disables itSame as MAI Code 2Admin must enable the policy before users can access itGitHub launch, GitHub June 26
Vendor performance claim98% of GPT 6.1 Sol's average across four benchmarksDescribed as comparable to two rival models, with no scoresSWE-bench Verified 72.6% in trade coverageMicrosoft AI, The Decoder
Independent check foundNone in our searchNone in our searchTrade coverage, Aug 2026The Decoder
Vision inputNot statedNot statedNot statedMicrosoft AI
Context windowNot statedNot statedLocal build keeps a 256K context windowMicrosoft AI 1.1
Where you can select itVS Code, Visual Studio, Copilot CLI, cloud agent, Copilot app, github.com, mobile, JetBrains, Xcode, EclipseSame list as MAI Code 2Copilot CLI, cloud agent, Copilot app, Copilot Chat on GitHub, Visual Studio, mobile, JetBrains, Eclipse, XcodeGitHub launch, GitHub June 18

The matrix shows the main split. MAI Code 2 is the only one of the three that excludes the Free and Student plans, and it is the only one whose headline claim is a multi-model average rather than a single named benchmark score. Microsoft's page says the averaged figures run across DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1, but it does not publish each benchmark score separately. That is a reason to ask for per-task evidence before you move your hardest work.

What it costs per month

The per-token rates below come from GitHub's billing docs. Promotional rates apply to the two new models, and the docs list the post-promotion rates in footnotes.

ModelInput per 1M tokensCached input per 1M tokensOutput per 1M tokensAfter promotion (input / output)
MAI Code 2$0.50$0.025$2.50$1.00 / $5.00
MAI Code 2 Flash$0.05$0.005$0.25$0.10 / $0.50
MAI-Code-1.1-Flash$0.20$0.02$1.20No promotional footnote listed

Pricing checked November 17, 2026.

Input rate for MAI Code 2 is $0.50 per 1M tokens according to GitHub Docs (2026). The docs state that 1 AI credit equals $0.01 USD, and the promotion runs through May 17, 2027 (GitHub's billing docs). On current promotional rates, the heavier model's input rate is 2.5 times the prior Flash model's, and its output rate is about 2.1 times higher. MAI Code 2 Flash's input rate is a quarter of the prior model's.

Plan prices sit on a separate page. GitHub's plans page lists Pro at $10 per user per month with $15 in monthly credits, Pro+ at $39 with $70, and Max at $100 with $200, according to GitHub (2026). The page shows no Student, Business or Enterprise price, but GitHub's plans documentation lists Business at $19 per seat per month with 1,900 monthly AI credits per user, and Enterprise at $39 per seat per month with 3,900.

PlanMonthly priceMonthly usage creditsMAI Code 2MAI Code 2 Flash
Free$0Limited chat and agent usageNoYes
Pro$10 per user per month$15YesYes
Pro+$39 per user per month$70YesYes
Max$100 per user per month$200YesYes
Business$19 per seat per month1,900 AI creditsYesYes
Enterprise$39 per seat per month3,900 AI creditsYesYes
StudentNot publishedNot statedNoYes

Here is an illustrative example, not a measured result. Suppose a team runs an automated review on every pull_request event, 200 times a month, and each run reads 300,000 input tokens and writes 20,000 output tokens. At the current promotional rates in the first table, MAI Code 2 costs 0.3 million × $0.50 = $0.15 for input plus 0.02 million × $2.50 = $0.05 for output, which is $0.20 a run and $40 a month. At the post-promotion rates it is $0.30 + $0.10 = $0.40 a run, or $80 a month. MAI Code 2 Flash comes to $0.015 + $0.005 = $0.02 a run, or $4 a month, and the prior MAI-Code-1.1-Flash to $0.06 + $0.024 = $0.084 a run, or $16.80 a month. These sums ignore cached-input discounts, retries and any plan credits, so use them to compare orders of magnitude rather than to forecast an invoice.

Vendor profiles: where each model fits

MAI Code 2

Best fit: Teams whose agent work is mostly multi-file changes, deep debugging and repository-wide reasoning, on Pro or higher, or on Business and Enterprise. Microsoft positions it for this work, so the fit claim is the vendor's, and it is worth testing on your hardest tickets before you rely on it.

Limitations: It is not listed for Free or Student plans, and it has the highest per-token list rate of the three models. Microsoft says the model is not tuned for "fundamental science research" or "very long horizon agentic work" (Microsoft AI).

Implementation: Confirm the plan tier for each seat, check the model policy for Business and Enterprise, and select the model in the picker for the work types above. Because cost scales with context length, set a review of token use after the first month.

Primary evidence: GitHub's launch changelog and Microsoft AI's announcement.

MAI Code 2 Flash

Best fit: Everyday explanations, targeted edits and routine fixes. It is listed for every Copilot plan, including Free and Student, so it is the cheapest way to let a whole team try the family. Microsoft says it is comparable to GPT-6 Luna and Claude Haiku 5.5 across three coding benchmarks it does not name, and it gives no scores.

Limitations: No published scores for Flash. Microsoft says Flash is in limited-access preview in Microsoft Foundry, which matters only if you plan to call the model outside Copilot. The page says the models will "soon" reach other platforms, and it does not give dates.

Implementation: Make Flash the default for routine work if your admin policy allows it, and set a written rule for when a task escalates to MAI Code 2.

Primary evidence: Microsoft AI's announcement and GitHub's launch changelog.

MAI-Code-1.1-Flash (the prior Flash model)

Best fit: Teams that want a mid-tier option or a local build. Microsoft says the model can be downloaded and run locally, recommends more than 120GB of RAM for best local performance, and says the local version keeps a full 256K context window (Microsoft AI 1.1).

Limitations: It is more expensive per token than MAI Code 2 Flash on list rates, and it requires an admin to enable the policy for Business and Enterprise users (GitHub June 26).

Implementation: If your team already uses it, compare its usage against the new Flash model before changing the default. Document which model each team member uses so that a switch does not leave two defaults in place.

Primary evidence: Microsoft AI's 1.1 page and GitHub's June 18 changelog.

Signal vs Speculation

Signal: published facts and figures.

  • Launch date, plan lists, admin defaults and billing mode come from GitHub's launch changelog and billing docs.

  • Per-token rates and the promotion end date come from GitHub's billing docs.

  • Benchmark averages and cost ratios come from Microsoft and are vendor-reported. We found no independent evaluation of MAI Code 2 in the sources we opened.

  • The Decoder's August 2026 coverage reports SWE-bench Verified at 72.6% for the 1.1 Flash model, and Microsoft's own 1.1 page claims 25% fewer tokens to complete a task.

  • Stack Overflow's 2025 survey gives the wider context: developers are adopting these tools, but they are not yet confident in them.

Speculation: our reading.

  • Microsoft's 98% average will matter less to your team than your own results on your own repositories. Treat the claim as a hypothesis to check.

  • Flash will absorb most routine work in teams that adopt it. That is a judgment about typical task mix, not a vendor fact.

  • The rename from the June model to the 1.1 billing name is probably a version bump. GitHub has not said so in the pages we opened.

  • The weights in the evaluation table are our judgment, not a standard.

Developers using or planning to use AI tools: 84% according to Stack Overflow (2025). The same survey finds that 46% of respondents actively distrust AI tool accuracy, against 33% who trust it, according to Stack Overflow (2025). That gap is why a vendor's benchmark average should not be the only input to a default-model decision.

MAI-Code-1.1-Flash SWE-bench Verified: 72.6% according to The Decoder (2026). That is the only performance number reported in trade coverage that we found for the MAI family in this search, and it is for the prior Flash model, not MAI Code 2.

Microsoft's MAI Code 2 page says its averaged score is 98% of GPT 6.1 Sol's and 96% of Claude Opus 5.5's, and that its cost is 40% of GPT 6.1 Sol's and 5% of Claude Opus 5.5's for that average, according to Microsoft AI (2026). Microsoft's 1.1 page claims the Flash model costs one quarter of the June model's price, according to Microsoft AI (2026).

Who this is for

  • Pick MAI Code 2 if your engineers spend most of their agent time on multi-file changes, and your team is on Pro or above, or on Business or Enterprise.

  • Pick MAI Code 2 Flash if you want a low-cost default for explanations and small edits, or if some users are on Free or Student plans.

  • Keep your current setup if your admin has not yet reviewed the default model policy, or if you need the local build Microsoft describes.

  • Skip the switch this month if every seat is on Free or Student and the main need is editor autocomplete.

Red flags: your work is mostly small edits, nobody has checked the admin model policy, or the budget owner needs a firm per-seat price before any usage-based model is used.

Where a workflow tool fits instead of a model swap

For many small teams, the real alternative is not a Copilot setting. It is stitching review or triage steps together in Zapier, Make or n8n, or writing a script in-house. Those routes can keep run histories, retry failed steps, branch on errors and keep audit records when they are configured that way. The buyer still designs and owns observability, idempotency (so a retried event does not post twice), escalation paths, access controls and ongoing maintenance. For how coding agents compare on scripted back-office work, see Claude Code vs OpenAI Codex for accounting automation scripts.

US Tech Automations could configure a review step as a proposed workflow. The trigger is a pull_request event. The workflow collects the diff and any linked issue, routes routine diffs to a Flash-class model and larger diffs to a heavier model, and writes a draft review comment to a queue. The output is that comment plus a log line with the model name, token count and run ID. Prerequisites are a GitHub token scoped to the repositories involved, a model API key with access to the chosen model, and confirmation that the model is reachable outside Copilot, since the pages we opened describe limited access for Flash in Foundry and no general API path for MAI Code 2. The human review point is an engineer who approves each draft before anything is posted. For how GitHub events connect to other tools, see how to connect Jira to GitHub automation.

A second proposed workflow handles the monthly cost check. On the first business day of each month, it would pull the Copilot usage export, group spend by repository and team, and flag any repository whose run count or token count rose sharply. The output is a short report to the operations lead. The prerequisite is a usage export with per-model detail, and the pages we opened do not describe such an export, so confirm that exists before designing around it. The human review point is the budget owner, who approves any change to model defaults.

When NOT to use US Tech Automations: if your need is only editor autocomplete, a Copilot seat on the default model is simpler than any workflow, so skip the build. If your only use is an occasional chat session, a manual prompt costs less time than configuring anything. If your repositories have no pull-request step worth automating, there is nothing for a workflow to trigger on, and the model picker is the whole decision.

Common mistakes

  • Treating MAI Code 2 Flash as a cheap copy of MAI Code 2. Microsoft and GitHub position them for different work.

  • Assuming a vendor's average benchmark score predicts results on your repositories.

  • Forgetting that the promotional rate ends, so budgets built on current rates will rise.

  • Missing the admin model policy on Business and Enterprise, which can leave users unable to select the model or can enable it without review.

  • Reading "annual plans" in the launch note as a general price or as a guarantee of availability.

  • Sending every task to the heaviest model, which raises cost on work the Flash model can handle.

Decision checklist for this month

  • Check each seat's plan tier against the plan table above.

  • On Business or Enterprise, confirm the model policy in Copilot settings before users are told which model to pick.

  • Pick one routine task and one multi-file task, run each through the model your admin allows, and record the token counts your usage view shows.

  • Work out your monthly ceiling with the worked-example arithmetic, using your own token counts instead of the illustrative ones.

  • Write down the default model and the rule for escalating to the heavier model, so the choice does not depend on individual habit.

Questions buyers ask first

Can Free and Student users select MAI Code 2?

No, GitHub's launch note lists MAI Code 2 for Pro, Pro+, Max, Business and Enterprise only, while MAI Code 2 Flash is listed for every plan including Free and Student.

Does MAI Code 2 replace MAI-Code-1.1-Flash?

The pages we opened do not say that it replaces the prior model, and GitHub's billing table lists the 1.1 name and no longer lists the June name, so check what your model picker shows.

Is MAI Code 2 cheaper than the prior Flash model?

No, on list rates the heavier model costs more per token than MAI-Code-1.1-Flash, while MAI Code 2 Flash costs less.

Do admins need to turn anything on?

For Business and Enterprise, check the model policy in Copilot settings, because GitHub says new models are enabled automatically unless an administrator turns off the global default or disables the model.

Is there independent proof of the performance claims?

Not in the sources we found, so the 98% average is Microsoft's own figure, and we found no third-party benchmark of MAI Code 2.

What to do this month

Start with the plan and the policy, not the benchmark. Confirm which seats can reach MAI Code 2, set Flash as the routine default if your admin allows it, and reserve the heavier model for multi-file work where its rate can be justified against the worked example. If the work needs a repeatable review step outside Copilot, see how US Tech Automations configures this as a proposed workflow with a human approval point before anything is posted.

About the Author

Garrett Mullins
Garrett Mullins
Workflow Specialist

Helping businesses leverage automation for operational efficiency.