Microsoft MAI Code 2 debuted at Ignite on Nov. 17, and the pitch to developers was less about beating frontier models outright and more about getting embarrassingly close for pocket change. The Microsoft MAI Code 2 release, alongside a lighter MAI Code 2 Flash, arrived as part of a broader slate of in-house MAI model drops positioned squarely on price-performance against OpenAI and Anthropic.
The headline claims, straight from Microsoft: MAI Code 2 reaches 98% of GPT-6.1 Sol’s performance at 40% of the cost, and 96% of Claude Opus 5.5’s performance at 5% of the cost. Those are vendor-reported numbers – the industry’s equivalent of a car ad’s “up to” mileage – but the benchmark set at least looks like the real work: DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1, averaging a 63.2% pass rate across the four.
Microsoft MAI Code 2 benchmarks and pricing
Microsoft positions MAI Code 2 for complex engineering: deep debugging, multi-file implementation, refactoring and repository-wide reasoning – the tasks that separate a coding assistant from a coding colleague. Availability runs through GitHub Copilot and Microsoft Foundry with usage-based billing.
In GitHub Copilot specifically, promotional pricing runs through May 17, 2027 at $0.50 per million input tokens and $2.50 per million output tokens, stepping up to $1.00/$5.00 standard afterward. It lands in Copilot Pro, Pro+, Max, Business and Enterprise plans.
MAI Code 2 Flash for everyday edits
Flash is the volume player: code explanations, targeted edits, routine fixes and rapid iteration, tuned across three benchmarks (FrontierCode v1.1 excluded) for a 61.0% average pass rate. Microsoft claims performance comparable to GPT-6 Luna at 65% of the price and Claude Haiku 5.5 at 20% of the price. Pricing undercuts sharply: $0.05/$0.25 per million tokens on promo, $0.10/$0.50 standard – and it’s available even on Copilot Free and Student tiers, which tells you exactly who Microsoft wants hooked.
Why the price-performance claims matter
Frontier-model comparisons age fast; the models Microsoft measured against may already be superseded. What’s structural is the direction of travel: in-house models claiming near-frontier quality at single-digit percentages of frontier cost, distributed through channels developers already pay for. Coding has been the AI sector’s most defensible paid use case – if the marginal model gets this cheap, the interesting question shifts from “can it code?” to “what do you build when inference is nearly free?” Microsoft would like the answer to run through its foundry.
