MAI Code 2 and Flash: Copilot Prices and Access
MAI Code 2 and Flash are in GitHub Copilot. Compare plan access, promotional token prices, and a practical way to check your real coding cost.

On November 17, 2026, Microsoft introduced MAI Code 2 and MAI Code 2 Flash, and GitHub added both to Copilot. Code 2 is available on Pro, Pro+, Max, Business and Enterprise; Flash also reaches Free and Student. GitHub’s Copilot pricing table lists promotional input prices of $0.50 and $0.05 per million tokens, respectively, through May 16, 2027. The published rates double on May 17. Check your plan, model policy and actual usage before choosing one for a coding workflow.
Pick the model for the task
Microsoft’s announcement describes Code 2 for deep debugging, changes across multiple files and reasoning over a repository. It describes Flash for code explanations, targeted edits, routine fixes and quick iterations. Those are the maker’s intended uses, not results of a test by DailyTech.
| Model | Work Microsoft and GitHub describe | Copilot plans GitHub lists |
|---|---|---|
| MAI Code 2 | Deep debugging, multi-file implementation, refactoring and repository-wide reasoning | Pro, Pro+, Max, Business, Enterprise |
| MAI Code 2 Flash | Explanations, targeted edits, routine fixes and rapid iteration | Free, Student, Pro, Pro+, Max, Business, Enterprise |
GitHub says the models expand its MAI family from lightweight coding assistance into more complex agentic software engineering. That describes the scope of this release; it does not establish that either model will outperform the model you already use on your own code. A small trial with representative tasks is more useful than selecting solely by the model label.
You can select either model in the Copilot model picker where your plan and client support it. GitHub names Visual Studio Code, Visual Studio, Copilot CLI, cloud agent, Copilot app, GitHub.com, GitHub Mobile, JetBrains, Xcode and Eclipse. It also says eligible subscribers on annual plans can access the models. For Business and Enterprise, an administrator can manage MAI access through model policy. Under default enablement, new models are enabled automatically unless the administrator has turned off that global default or disabled the particular model. A missing picker entry therefore calls for a plan, client and policy check before a performance judgment.
Read the promotional prices carefully
GitHub bills these models under usage-based billing. Its current model pricing reference states the following US dollar rates per one million tokens. Input, cached input and output have separate prices; the promotional period ends before May 17, 2027.
| Model | Period | Input / 1M | Cached input / 1M | Output / 1M |
|---|---|---|---|---|
| MAI Code 2 | Through May 16, 2027 | $0.50 | $0.025 | $2.50 |
| MAI Code 2 | From May 17, 2027 | $1.00 | $0.05 | $5.00 |
| MAI Code 2 Flash | Through May 16, 2027 | $0.05 | $0.005 | $0.25 |
| MAI Code 2 Flash | From May 17, 2027 | $0.10 | $0.01 | $0.50 |
These are Copilot rates, not a promise about a separate API or Microsoft Foundry bill. Microsoft says Flash is in a limited-access Foundry preview. It does not say that Copilot plan eligibility grants Foundry access. GitHub converts token charges into AI credits, with one credit worth $0.01. Included credit allowances depend on the plan, and usage beyond those allowances can be billed according to the account’s budget settings. People on legacy annual request-based plans should check GitHub’s separate model-multiplier rules before treating the token table as their own billing method.
For scale, an illustrative interaction with one million uncached input tokens and 100,000 output tokens would cost $0.75 on Code 2 at the promotional rates: $0.50 input plus $0.25 output. The same token counts would cost $0.075 on Flash: $0.05 plus $0.025. That arithmetic excludes cached tokens, retries, tool calls and any plan allowance. It is not a bill we observed. Use the API cost calculator to explore token arithmetic, then use GitHub’s Copilot usage record and pricing table for the charge that applies to your account.
What Microsoft’s performance comparison measures
Microsoft says Code 2 reached 98% of GPT-6.1 Sol’s performance at 40% of its cost, averaged across four named coding benchmarks. It also reports 96% of Claude Opus 5.5’s performance at 5% of its cost. These are Microsoft’s own comparisons, not independently reproduced results. Its footnotes say the models ran at maximum reasoning effort, three benchmark sets ran on Microsoft’s internal evaluation platform through Copilot CLI without subagents, and the Code 2 cost used its 50% promotional price. The Flash comparison also uses the promotional price. Microsoft warns that workload, configuration, reasoning effort and pricing assumptions change the result.
The published ratios cannot tell you the cost of an accepted change in your repository. A task that needs several retries, expensive review or a larger context can change the total even when a per-token rate is low. The Copilot plan comparison explains how subscription allowances differ from usage charges. If your team is also replacing a retiring model, the December retirement checklist covers that separate access decision.
Run a small, recorded trial
Choose a routine edit and a difficult multi-file task you already know how to review. Run each with a model your account permits. Keep the prompt, repository state and acceptance test as similar as practical. Record the following for each run:
| Record | Why it matters |
|---|---|
| Task, model, Copilot client and plan | Makes the comparison repeatable and shows where access applies. |
| Token or AI-credit spend, including retries | Captures the observed cost rather than the headline rate alone. |
| Whether the change was accepted | Separates useful output from a response that still needs rework. |
| Human review time and failed checks | Exposes work that a cheap model response can shift to a maintainer. |
For a group of comparable tasks, divide total observed spend by the number of accepted changes. If none is accepted, report that no successful change was produced; there is no valid cost-per-accepted-change denominator. Compare review time alongside the spend. This worksheet is a suggested method, not a DailyTech test or a claim that one model is better. The AI model tracker gives broader release context; your Copilot picker, billing record and reviewed changes decide the choice here.
Sources
- Microsoft AI, November 17 announcement — intended workloads, benchmark claims, configurations and promotional-price footnotes.
- GitHub Changelog, November 17 — launch, plan and client availability, billing type and administrator policy.
- GitHub Docs, Models and pricing for GitHub Copilot — token rates, promotion end date, credit conversion and annual-plan caveat.
All pages read on November 18, 2026.
Free toolWhat will 10,000 chats a month cost?Estimate my bill

