Two models, two usage tiers
During the Ignite 2026 event, Microsoft announced MAI Code 2 and MAI Code 2 Flash, two programming models developed in-house and made available on GitHub Copilot. According to the source’s account, the two products cover different needs: Code 2 is designed for complex software development, such as in-depth bug detection, changes spread across multiple files, understanding an entire repository, and building user interfaces; Code 2 Flash is instead geared toward light and repetitive tasks, from code explanation to targeted fixes and rapid iterations.
According to the announcement, both models support working modes in which the AI calls tools and proceeds through successive steps while carrying out a development task. This is a description of the capabilities claimed: the material does not offer independent evidence to assess how well these modes work in practice.
The per-task cost lever
Pricing is the main launch argument. Microsoft states that, on average across four programming benchmarks published by the company itself, Code 2 reaches about 98% of GPT-6.1 Sol’s performance at a cost equal to about 40% of the latter’s. The cited data indicate an average cost per task of $0.309 for MAI Code 2, versus $0.761 for GPT-6.1 Sol and $6.306 for Claude Opus 5.5: compared with the latter, Code 2 would therefore cost about 5%.
For the Flash version, with a 50% promotion, Microsoft reports an average cost per task of $0.052, equal to 65% of GPT-6 Luna and 20% of Claude Haiku 5.5. On performance, the company reports an average pass rate of 61.0% on DeepSWE 1.1, Terminal-Bench 2.1, and SWE Atlas, versus 55.1% for GPT-6 Luna and 53.9% for Claude Haiku 5.5.
The key point is the comparison model adopted: not only the quality of the code produced, but the cost incurred for each individual task performed. If the data were confirmed, the indicated consequence would be a reduction in spending on AI use in development, especially for companies that make many calls. It should be noted, however, that all the numbers come from tests and communications by Microsoft itself and cannot be verified with the information available.
Access and open questions
In terms of distribution, Code 2 is reserved for users with eligible Copilot Pro, Business, and Enterprise subscriptions, while Flash also includes Free and Student users. The choice allows companies to assign the more demanding tasks to the larger model and repetitive ones to the cheaper model, modulating spending according to workload.
Some aspects that the source does not address remain to be clarified: the exact composition of the four benchmarks cited, the method for calculating the cost per individual task, the duration of Flash’s promotional price, and any impact on already active plans. In the absence of these elements, the actual scope of the announced savings and the comparison with competing models should be considered as company statements pending independent confirmation.
