AI NewsModels & agentsAnnouncement
Microsoft released MAI Code 2 and MAI Code 2 Flash
Microsoft released two coding models at Ignite. MAI Code 2 is for hard work. MAI Code 2 Flash is for fast edits.

Why it mattersA team paying for Claude or GPT for coding has a cheaper third option to test. Every speed and cost number is Microsoft's own.
A team paying for Claude Opus on code review gets a bill at the end of the month. Microsoft says its new coding model does 96 percent of that work for 5 percent of the cost. Microsoft released MAI Code 2 and MAI Code 2 Flash on Tuesday at its Ignite conference. Both models are available in GitHub Copilot. MAI Code 2 is also in a limited-access preview in Microsoft Foundry. Every speed and cost number here is Microsoft's own.
MAI Code 2 is the model for harder work. Microsoft says it is for moving code to production, deep debugging, multi-file changes, whole-repo reasoning and UI work. MAI Code 2 Flash is for routine jobs. The post names code explanations, small edits, routine fixes and quick iterations. Microsoft announced both models at its Ignite conference in San Francisco.
Microsoft's own benchmark numbers
Microsoft says MAI Code 2 scores at 98 percent of GPT-6.1 Sol at 40 percent of the cost. It says MAI Code 2 scores at 96 percent of Claude Opus 5.5 at 5 percent of the cost. The scores are an average across four tests: DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1. Microsoft also ran a front-end review. It says 258 annotators preferred MAI Code 2's output over GPT-6 Sol's output 67 percent of the time. For MAI Code 2 Flash, Microsoft says the model matches GPT-6 Luna at 65 percent of the cost. It matches Claude Haiku 5.5 at 20 percent of the cost. These are Microsoft's own tests.
A reader should treat every one of those figures as the vendor's. Microsoft used an internal benchmark set it calls MSBench. The cost figures are sourced from Artificial Analysis. An independent third party has not scored either model yet. A buyer swayed by "Opus performance at 5 percent of the cost" should also note that the comparison is Microsoft's.
How to try it
Both models are available in GitHub Copilot today. MAI Code 2 is in limited-access preview in Microsoft Foundry. Microsoft says more hosts will offer the models soon. The post names Open Router, Vercel, Hermes, Open Code and Kilo. Microsoft listed a 50 percent promotional price in the post. The full price is not stated.
What a buying team should do
The useful check is one Reveneau runs with clients. Pick one real pull request from last week's codebase. Run the current model and MAI Code 2 Flash on the same prompt. Grade both against the pull request's final state. Count the tokens each used. Then run the harder test on one bug that took a senior two days to debug. Score with the team's own evals, not Microsoft's.
The point under the price gap is that top-end coding models now come at many price points. A team can send easy work to a Flash-class model and harder work to a reasoning model. This two-model setup has quietly become normal in AI engineering. A new cheap model at the top end makes the setup save money sooner. The rule still holds. The vendor picks the benchmark. The buyer picks the real test. The buyer's test is the one that matches the real bill.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.


