Microsoft introduced MAI Code 2 and MAI Code 2 Flash on November 17, making both coding models available in GitHub Copilot. Code 2 Flash is also entering a limited-access preview in Microsoft Foundry. The launch pairs different engineering workloads with promotional token prices scheduled to end on May 17, 2027.
Two models for different engineering workloads
Microsoft positions Code 2 for work across existing codebases, including debugging, changes spanning several files, repository-level reasoning and interface development. Flash targets high-volume tasks such as explaining code, making focused edits and iterating on routine fixes.
The distinction gives teams another choice within the GitHub Copilot model lineup7.3 out of 10. It also makes task selection relevant to cost: a short explanation and a complex refactor need not use the same model. Microsoft’s positioning describes the intended use of each model; it does not establish that either will outperform alternatives on a team’s own repository.
Promotional rates have a published end date
GitHub lists Code 2 at $0.50 per million input tokens, $0.025 per million cached input tokens and $2.50 per million output tokens. Flash costs $0.05, $0.005 and $0.25 respectively.
Those are promotional rates. From May 17, 2027, GitHub lists Code 2 at $1 for input, $0.05 for cached input and $5 for output. Flash moves to $0.10, $0.01 and $0.50. All figures are per million tokens.
Copilot converts token consumption into AI credits, with allowances and billing depending on the plan. These rates therefore describe model usage inside Copilot, rather than a standalone subscription price or a promise about the total cost of an agent task.
Benchmark comparisons use the discounted prices
Across four coding benchmarks, Microsoft reports Code 2 at 98% of GPT-6.1 Sol’s performance for 40% of the cost. Its methodology uses maximum reasoning effort and includes a 50% promotional price for Code 2. Several evaluations ran on Microsoft’s internal platform; FrontierCode figures came from Artificial Analysis.
Microsoft explicitly says the family is not yet tuned for fundamental scientific research or very long-horizon agentic work. That narrows the claim: these are models aimed at everyday software engineering, with launch comparisons tied to specific tasks, settings and temporary prices.