Microsoft brings its MAI Code 2 models to GitHub Copilot at a temporary discount
Microsoft has made two of its own programming models available in GitHub Copilot: MAI Code 2 and MAI Code 2 Flash. The introductory price will double on 17 May 2027.
, Microsoft released two of its own programming models in GitHub Copilot, the coding assistant of its subsidiary GitHub: MAI Code 2 and the lighter MAI Code 2 Flash. Until 17 May 2027 they cost half the normal price. This makes it attractive to choose Microsoft's own models instead of those of other makers, such as OpenAI and Anthropic, which can also be chosen in Copilot.
MAI Code 2 is meant for heavier programming work, such as tracking down difficult bugs, changing several files at once and keeping an overview of all the code in a project. The Flash version focuses on everyday work, such as explaining code, small targeted changes and simple fixes. Both are made for work in which an AI agent carries out programming tasks by itself. According to Microsoft, they are not yet tuned for harder tasks, such as fundamental scientific research or agent work that runs for a very long time without a break.
Performance and price
According to Microsoft, MAI Code 2 reached 98 per cent of the score of OpenAI's GPT-6.1 Sol, averaged over four widely used tests for programming models, at 40 per cent of the cost. Compared with Anthropic's Claude Opus 5.5, it reached 96 per cent of the score at 5 per cent of the cost. On three tests for routine work, Code 2 Flash performed about as well as the lighter models GPT-6 Luna and Claude Haiku 5.5, at 65 and 20 per cent of their cost.
These comparisons use the half-price introductory rate. Three of the tests ran on Microsoft's own test system; the scores and costs of the fourth come from the independent testing firm Artificial Analysis. Results can differ per task and setting.
In another test, 258 people judged screen designs, games, charts and 3D scenes without knowing which model had made what. They rated the work of MAI Code 2 as better than that of GPT-6 Sol in 67 per cent of cases.
Users pay per use. Until 17 May 2027, MAI Code 2 costs 0.50 dollars per million tokens (small pieces of text) that the model reads and 2.50 dollars per million tokens that it writes; after that, these prices double. Flash costs a tenth of that.
Availability
Code 2 can be chosen in the paid Copilot plans Pro, Pro+, Max, Business and Enterprise; Flash is also available in the free version and the version for students. According to GitHub, this applies to subscribers with an annual plan. Microsoft and GitHub do not mention separate rules for Europe. Flash can also be used in a limited test version through Microsoft Foundry, Microsoft's AI platform for businesses. An adapted version, MAI Work Flash, already works in Excel, PowerPoint, Word and Teams.
What does this mean for the road to the singularity?
Developers could already choose between models from different makers in Copilot. What is new is that Microsoft, which runs Copilot itself, now puts its own models on the list, and according to the company they are almost as good for a fraction of the price. For OpenAI and Anthropic, that creates pressure to lower their prices for ordinary programming work, or to show what more their more expensive models can do.
If that price pressure continues, an AI coding assistant for everyday work will become so cheap that small businesses and hobbyists can also let it help all day long. The most expensive models will then be left for the hardest work, such as scientific research, for which Microsoft itself says it is not ready yet. Anyone who wants to win there has to aim their money and computing power at that frontier, and that is exactly where AI can move forward the most.
What is the technological singularity? Read the explanation →