Microsoft’s MAI Code 2 Flash Is Here. Is It Worth It?
I found out I was using MAI Code 2 Flash after I had already spent time correcting its work. I was working on a serious project in GitHub Copilot. I asked for changes, explained what I wanted, and told the assistant what not to do. It kept making mistakes I had already pointed out. Eventually, I wrote part of the code myself and gave it to the model to follow.
Then I checked which model Copilot was using. It was MAI. I didn’t remember choosing it, so I switched to Claude Sonnet. Within a few minutes, Claude understood the change I wanted and fixed work that MAI had left unfinished. Claude also cleaned up the entire mess MAI left behind.
That experience matters more to me than a benchmark score. Here’s what Microsoft says about the model, what another developer found, and why I wouldn’t choose Flash again for this kind of work.
What is claimed
Microsoft released MAI Code 2 and MAI Code 2 Flash on November 17. These are two different models. This article is about Flash, the one I used in Copilot. Some of the strongest claims in the announcement are about the larger MAI Code 2. They shouldn’t be used to judge Flash, just as my experience with Flash shouldn’t be used to judge the larger model.
GitHub describes Flash as a good choice for explaining code, making targeted edits, fixing routine problems, and working through changes quickly. Microsoft says that, across three coding benchmarks, Flash performed about as well as GPT-6 Luna and Claude Haiku 5.5, at 65% and 20% of their respective costs.
I understand why that sounds attractive. If two tools can do the same job and one costs less, there’s a good reason to use the cheaper one. But the important part is doing the same job. A lower price tells me what it costs to get an answer. It doesn’t tell me how much work I’ll have to do after I get it.
There’s also a distinction worth making: Microsoft compared Flash with Claude Haiku in those figures. I switched from Flash to Claude Sonnet. I’m not claiming my experience disproves Microsoft’s numbers. I’m saying those numbers don’t answer the question that mattered to me: which model could follow my directions and help me finish my work?
What people say
One account I found useful was Jake Groszewski’s review of both MAI models. He tried them on several Python tasks and described where Flash helped and where it fell short.
When he asked Flash to explain some data-processing code, it gave him a clear answer and noticed what would happen if the input contained no data. The larger model took longer and gave him more detail than he needed. For that task, he preferred Flash. That’s a good result: the faster model answered the question, caught something useful, and let him get back to work.
He also asked for a change that involved several parts of an application. Flash changed the screen and the code that responded to user actions, but it skipped the changes needed to store the data. He reported having to follow up twice to get the complete implementation.
Those are two findings from one developer, not a verdict from everyone who has tried MAI. What makes the review useful is that it explains why the experience differed. When you ask for an explanation, you can read the answer and decide whether it helped. When you ask for changes to a project, you have to check not only what the assistant changed but also what it left untouched. A fast answer saves less time if you’re the one who has to find the missing work.
What was my experience?
I wasn’t trying out prompts to see which model would win. I had work to do. I’m keeping the details of the project private, but one exchange shows the problem clearly enough.
I pointed out a piece of code that was in the wrong place. In the same message, I said there were many other pieces of code there with the same problem. MAI searched the relevant part of the project, moved the one I had named, and reported that the build and a test had passed. The others were still there.
I had to tell it again that I meant the others too. That wasn’t a new requirement I had thought of after seeing the result. It was in my first message.
What bothered me wasn’t simply that the model missed something. It was that its reply made the work sound finished. It could tell me exactly what it had moved and that the code had built successfully. Both details sounded reassuring, but neither answered the question I actually cared about: had it dealt with the whole problem? To find that out, I had to go back and check for myself.
My frustration with MAI went beyond that exchange. I would tell it not to do something, see the same mistake again, and have to correct it again. This happened more than four or five times. MAI kept reverting my changes and reapplying the exact changes I had told it not to make. There’s a difference between spending time explaining a project and spending time repeating a decision you have already made. The first is part of working with someone. The second is what made me feel I was wasting my time.
It also changed how much I could trust the next edit. Once an assistant repeats a mistake I’ve already corrected, I can’t just review the new work. I also have to look for the old mistake. And if it says everything is done, I still need to check whether “done” means the change I requested or just the part it happened to make.
That’s how I ended up writing code manually and giving it to MAI to follow. I wasn’t giving a demonstration or running a test. I needed a particular result in my project, and my instructions weren’t getting me there. By then, the cost wasn’t just the time spent writing that code. It was also the time I had already spent asking, reading the results, finding the problems, and correcting them.
The model selection added to my frustration. I was using Copilot. I didn’t remember selecting MAI. Who selected MAI Code 2 Flash? I switched to Claude Sonnet, and it handled the request. What I do know is that I had already lost trust in Copilot before I chose a model to work with.
When I switched to Claude Sonnet, I didn’t write a new specification. I repeated the point that the change needed to cover the rest of the affected code. Less than four minutes later, Claude had fixed all my problems and gone even further. The difference was easy to see in the work itself. Before the switch, MAI kept going against my instructions, putting functions in the wrong files and making a mess despite my repeated explanations. After the switch, Claude reorganized the related code too.
Claude was good. It understood what I wanted and helped fix the mess. That mattered to me because I could finally spend my time reviewing the change I had asked for, rather than explaining once again why the previous change was incomplete.
Conclusion
I don’t need to argue that Microsoft’s benchmarks are wrong or that someone else had a bad experience when they say they had a good one. Flash may be fast and useful for explaining a piece of code. My problem was different: I needed an assistant that would follow my instructions, remember my corrections, and carry a change through a real project. Too much of that work came back to me.
The price that matters to me includes my own time. If I have to give the same direction several times, inspect each attempt for mistakes I’ve already explained, and finally write the code myself, a cheaper model hasn’t made the job cheaper for me.
What settled it wasn’t a score or an announcement. It was switching models in the middle of the work. The project hadn’t changed. Neither had the instruction. With Claude Sonnet, I got the change I had been asking for within minutes. For serious work in Copilot, I would rather choose that model myself than leave it unwatched.