At Microsoft Ignite on November 17, 2026, Microsoft announced specialized MAI models, including two coding models aimed at different kinds of work. The company’s performance and price comparisons for MAI Code 2 and MAI Code 2 Flash cover named coding benchmarks; GitHub separately listed which Copilot plans can access each model and the billing conditions. (Microsoft’s MAI Code announcement; GitHub’s Copilot changelog)
Practical read: Match each comparison to the coding work and benchmarks Microsoft names before using its percentages to judge a model for your own workload.
What do Microsoft’s coding-model comparisons cover?
Microsoft positions MAI Code 2 for everyday enterprise coding and MAI Code 2 Flash for routine, high-volume coding. The company reports Code 2’s results as averages across four benchmarks. For Flash, it describes performance as comparable across three routine-coding benchmarks, without giving a single percentage score for either comparison. (Microsoft’s benchmark details)
| Microsoft model | Comparator | Performance Microsoft reports | Relative price Microsoft reports | Benchmark scope |
| MAI Code 2 | GPT-6.1 Sol | 98% of the comparator’s performance | About 40% of its price | Average across DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1; everyday enterprise coding |
| MAI Code 2 | Claude Opus 5.5 | 96% of the comparator’s performance | 5% of its price | Average across DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1; everyday enterprise coding |
| MAI Code 2 Flash | GPT-6 Luna | Comparable performance | 65% of its price | DeepSWE 1.1, Terminal Bench 2.1 and SWE-Atlas; routine coding |
| MAI Code 2 Flash | Claude Haiku 5.5 | Comparable performance | 20% of its price | DeepSWE 1.1, Terminal Bench 2.1 and SWE-Atlas; routine coding |
The percentages and price ratios are Microsoft’s reported comparisons, tied to those benchmarks and task descriptions. Microsoft says its coding models are not yet tuned for fundamental science research or very long-horizon agentic work. That boundary matters if your work depends on those tasks: the everyday- and routine-coding comparisons do not answer how the models perform on them.
Which GitHub Copilot plans list each model?
GitHub’s November 17 changelog listed MAI Code 2 for Copilot Pro, Pro+, Max, Business and Enterprise. It listed MAI Code 2 Flash for those plans as well as Free and Student. GitHub specified eligible annual-plan subscriptions for both and usage-based billing. (GitHub’s plan and billing details)
| Model | Copilot plans GitHub lists | Subscription condition | Billing |
| MAI Code 2 | Pro, Pro+, Max, Business, Enterprise | Eligible annual-plan subscription | Usage-based |
| MAI Code 2 Flash | Free, Student, Pro, Pro+, Max, Business, Enterprise | Eligible annual-plan subscription | Usage-based |
For Business and Enterprise, GitHub says administrators can manage model access through Copilot settings. Usage-based billing means the plan list alone does not tell an individual subscriber what their total use will cost.
How does Microsoft’s enterprise customization offer work?
Microsoft’s MAI Licensed Weights offer lets enterprise customers customize an MAI model with proprietary data and run the customized model in their own Azure environment. (Microsoft’s Ignite announcement)