Microsoft’s MAI Code 2 arrives with low introductory token prices, but its performance comparisons depend on four benchmarks and a temporary discount.
Microsoft has turned the coding-model contest into a more direct argument about cost. Its MAI Code 2 and MAI Code 2 Flash models are offered at low introductory token prices, while the company says Code 2 approaches the performance of two leading rivals on a set of coding benchmarks. For software teams and small businesses, the useful question is not which model wins a headline comparison, but how well it handles their code at the price they will actually pay.
Microsoft announced MAI Code 2 on November 18. Across four coding benchmarks, the company says the model delivers 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. The cost comparisons use a 50% promotional price. That qualification matters: benchmark averages do not predict every repository, and a discount is not a permanent operating cost.
The published launch rates, reported by EveryDev.ai, are $0.50 per million input tokens and $2.50 per million output tokens for MAI Code 2. The lighter MAI Code 2 Flash is priced at $0.05 per million input tokens and $0.25 per million output tokens. On May 17, 2027, those rates are scheduled to rise to $1 and $5 for Code 2, and $0.10 and $0.50 for Flash. The price schedule gives developers a concrete reason to measure usage now and budget against the later rates, rather than treating the introductory offer as the long-term baseline.
The split between the models also suggests a practical starting point. A roundup from EveryDev.ai describes Code 2 as aimed at repository understanding, debugging and changes across multiple files, with Flash intended for explanations, focused edits and routine fixes. Those are sensible categories for a trial: route repetitive, lower-risk work to the cheaper model, then check whether Code 2 materially improves results on broader tasks. Any savings should be weighed against review time, failed edits and the cost of rerunning prompts.
The announcement does not establish that either model is best for a particular team. Its comparisons cover four benchmarks, not every language, framework or private codebase. Developers should run a small evaluation against representative tasks, compare output quality and completion rates, and track both input and output tokens. Keeping a fallback model available also reduces the risk of building a workflow around one vendor’s pricing or availability.
The week’s developer news roundup also described Google’s Gemini 4 Neon and a Google preview for retrieving tool definitions only when an agent needs them. Those claims are not sufficiently established in the available material to support firm conclusions here. In particular, the reported pricing and product details should be checked against Google’s own documentation before teams make procurement or architecture decisions.
GitHub Agentic Workflows and Vercel’s developer tools also appeared in the roundup, but the supplied information does not provide enough independently confirmed detail to assess their current availability or capabilities. For now, the clearest actionable change is Microsoft’s published price schedule. Builders can try MAI Code 2 and Flash on non-sensitive, representative tasks, compare them with their existing tools, and record the results before committing production work.
That is less dramatic than a claim that one model has won coding. It is more useful. As AI features become operating costs for small software businesses, transparent rates and repeatable testing matter more than a single benchmark score. Microsoft has put numbers on the table; the next step for buyers is to verify what those numbers mean in their own systems.

