On this page

On November 17, 2026 (GitHub Changelog JSON-LD published at 2026-11-17T18:01:45+00:00, or November 18, 2026 at 02:01:45 Beijing time, with visible page date November 17, 2026; Microsoft AI intro page verified via curl and Jina with visible date Nov 17, 2026 and JSON-LD timestamp November 18, 2026 at 02:15:16 Beijing time; Model Card PDF header dated 17 November 2026), Microsoft and GitHub launched MAI Code 2 and MAI Code 2 Flash in GitHub Copilot.

For developers using usage-based token billing in GitHub Copilot (1 AI credit = $0.01 USD), the two models establish a clear two-tier workflow. For seat subscription tiers, AI Credits rules, and earlier model tables, see our GitHub Copilot pricing guide, GitHub Copilot AI Credits explainer, earlier GitHub Copilot models list, and GitHub Copilot hub page (and note that these coding models are separate from MAI Cyber 1.1 Flash).

Official Positioning and Workload Fit

**When to Choose MAI Code 2: **Across the Microsoft AI announcement and GitHub Changelog, MAI Code 2 is positioned as a general-purpose coding workhorse suited for taking code to production, deep debugging, multi-file implementation, refactoring, repository-wide reasoning, and UI work. The Model Card summary and Section 1 likewise recommend MAI Code 2 for the most complex, long-horizon tasks.

**When to Choose MAI Code 2 Flash: **MAI Code 2 Flash is a lightweight, cost-efficient model built for high-volume workflows such as code explanations, targeted edits, routine fixes, and rapid iterations. The Model Card notes it is optimized for low latency and low serving cost, is less capable than MAI Code 2 on complex, long-horizon tasks, and may require more explicit guidance.

**Nuance Across Official Sources on Long-Horizon Tasks: **Official pages present slightly different framing on agentic scope: the GitHub Changelog states that both models are designed for agentic coding workflows and the Model Card recommends MAI Code 2 for complex, long-horizon tasks, while the Microsoft AI intro page notes that both models are "explicitly not yet tuned for harder problems, like fundamental science research and very long horizon agentic work."

Figure 1 Official Microsoft AI header visual introducing MAI Code 2 and MAI Code 2 Flash
Figure 1 Official Microsoft AI header visual introducing MAI Code 2 and MAI Code 2 Flash

Image source: Microsoft AI announcement page (introducing-mai-code-2-and-mai-code-2-flash).

Copilot Plan Access, Clients, and Distribution Channels

GitHub Copilot Plan Coverage: **Per the GitHub Changelog, **MAI Code 2 is available on Copilot Pro, Pro+, Max, Business, and Enterprise (Free and Student are not listed), whereas MAI Code 2 Flash is available on Copilot Free, Student, Pro, Pro+, Max, Business, and Enterprise. The Changelog also states that both models are available to eligible subscribers on annual plans, while the GitHub Docs Plans page notes that Free and Student plans access models through auto model selection only (without listing Free/Student columns in its model matrix). For Business and Enterprise organizations, both models are enabled automatically under default model enablement unless turned off by an administrator.

Supported Clients and API Channels: **Both models appear in the model picker across Visual Studio Code, Visual Studio, Copilot CLI, GitHub Copilot cloud agent, GitHub Copilot app, github.com, GitHub Mobile (iOS and Android), JetBrains, Xcode, and Eclipse. Outside GitHub Copilot, Model Card Sections 2.3 and 2.4 state that both models are hosted services without downloadable local weights, and that **MAI Code 2 Flash is available in limited-access preview in Microsoft Foundry via an OpenAI-compatible Responses API (public Foundry token rates and signup links are not published). The intro page also notes upcoming releases on Open Router, Vercel, Hermes, Open Code, and Kilo.

Official GitHub Copilot Token Prices and the May 17, 2027 Promo Window

According to GitHub Docs (models-and-pricing.md; all prices per 1 million tokens, with no separate Long context tier in the Microsoft table), both models launch at 50% promotional pricing until May 17, 2027, before moving to standard list rates starting May 17, 2027 (no timezone or hour is specified in either official footnote):

MAI Code 2 (Category: Powerful): **Promotional pricing until May 17, 2027 is **$0.50 per 1M input tokens, $0.025 per 1M cached input tokens, and $2.50 per 1M output tokens. Starting May 17, 2027, standard pricing is $1.00 input, $0.05 cached input, and $5.00 output per 1M tokens.

MAI Code 2 Flash (Category: Lightweight): **Promotional pricing until May 17, 2027 is **$0.05 per 1M input tokens, $0.005 per 1M cached input tokens, and $0.25 per 1M output tokens. Starting May 17, 2027, standard pricing is $0.10 input, $0.01 cached input, and $0.50 output per 1M tokens—exactly one-tenth of MAI Code 2's per-token rates across both periods (the older MAI-Code-1.1-Flash remains listed separately at $0.20 / $0.02 / $1.20 without a May 17, 2027 promo footnote).

Shared Architecture Specs and Official Benchmark Figures

Context Window and Reasoning Effort Levels: **Per the Model Card, both models were pre-trained and post-trained from scratch by Microsoft as decoder-only sparse MoE transformers with an integrated vision encoder, accepting text and image inputs (up to 922,000 input tokens) and producing text outputs (up to 128,000 output tokens) within a **1,050,000-token context length (also described as a 1M token context extension), with an August 2026 pre-training cut-off and English-optimized code generation at release. Both support configurable reasoning effort (low, medium, high, xhigh, max), though Section 2.7 item 6 cautions that higher effort increases latency and token usage and does not always improve results. For parameter counts, the Model Card summary table lists MAI Code 2 at 50B-100B (80B active) and Flash at 5B-15B (12B active), whereas Section 1 cites approximately 2.8 trillion total parameters (77 billion active, 3 billion embedded) for Code 2 and approximately 398 billion total (12 billion active) for Flash.

Official Benchmark Results: **In Model Card Section 3 (mean single-attempt pass rate across three runs on MSBench using GitHub Copilot CLI 1.0.102 without subagents at max reasoning effort), MAI Code 2 scores **73.2% vs. Flash's 59.9% on DeepSWE 1.1, and 84.6% vs. 78.3% on TerminalBench 2.1, while Flash posts 44.8% vs. Code 2's 43.1% on SWE Atlas. Across four benchmarks including Artificial Analysis's FrontierCode v1.1 (at max reasoning effort and 50% promotional pricing), Microsoft's intro page reports that MAI Code 2 achieves 98% of GPT 6.1 Sol performance at 40% of the cost, 96% of Claude Opus 5.5 performance at 5% of the cost, and a 64% preference rate over GPT-6 Sol in blind front-end UI evaluations, while MAI Code 2 Flash delivers comparable performance to GPT-6 Luna and Claude Haiku 5.5 at 65% and 20% of the cost (the MAI Code 2 model page separately cites 1/5 of Claude Haiku 5.5 cost with 12% higher performance).