Transcribing an equation from a paper into LaTeX by hand is one of those tasks that is not hard, takes four minutes, and destroys concentration every single time. Mathpix removes it. Snip the equation, get correct LaTeX, paste it. That is the whole product and the fact that there is nothing else to it is a strength rather than a limitation.
The documentation reflects that discipline, being precise, technical, short where it can be, and entirely free of the language most AI companies use. Nobody is transforming anything, they are converting an image to LaTeX at a stated accuracy. Accuracy on printed mathematics is excellent. Multi line equations, matrices, integrals, unusual notation, subscripts stacked in awkward ways, it handles them.
Handwriting is genuinely surprising, and while it is not perfect it is far better than a reasonable person would predict, which makes photographing a lecture board a workable input method. The output format range is well judged too, since LaTeX, Markdown and structured document formats are what technical users actually need, and the document conversion path for whole PDFs is more capable than the snipping feature it is known for. Where it struggles is layout complexity. A dense paper with multiple columns, floating figures, footnotes and tables will need manual repair after conversion, and the amount of repair scales badly with how ugly the original typesetting was.
The documentation is reasonably honest about this without giving a good sense of when to bother, and a practical page on which documents convert cleanly and which are faster to retype would save people time. The gap that matters most is verification. When you snip an equation from a field you understand, you notice if a subscript is wrong. When you are converting material you cannot read fluently, which is common for students and for anyone working across disciplines, an error passes straight through and lands in your notes or your paper.
The material could say much more about checking discipline, particularly on handwriting where the error rate is meaningfully higher and the errors are plausible rather than obvious. Chemistry and more specialised scientific notation are weaker than mathematics, which is understandable given where the effort went and worth knowing before you build a workflow around it. The free tier is tight enough that anyone with a real workload will pay quickly, and the pricing is fair for what it removes. Four point zero.
The clearest example on this site of a tool that picked one narrow problem, solved it properly, documented it honestly and resisted every temptation to become a platform. The only real criticism is that it could teach verification better, and that is a small complaint against a tool that does exactly what it claims.