How 15 AI Models Handle the Search for the Best XML Parser for iOS: Benchmark Results

A developer ran an unusual experiment: he fed the same task to 15 popular AI models and compared their results with his own manual benchmark of XML parsers for iOS.

What the task was

Six months ago, the author published the results of his own search for the fastest XML parser for iOS, tvOS and macOS. It was painstaking work: several hours of manually analyzing GitHub repositories, checking popularity (at least 500 stars), support for Objective-C and Swift, and integration via CocoaPods or SwiftPM. Three hours of hard work (and several cups of coffee) produced a full ranking of the best parsers.

Can AI do it faster

Then came the obvious thought: why spend 3 hours if the internet promises that AI can do it in 5 minutes? Moreover, there was a real chance that an error had crept into the manual benchmark, such as a misread piece of code or a missed detail in the specs. And if so, AI models with their enormous stores of knowledge might find a more accurate result.

The decision was natural: give the same task to 15 different AI models (OpenAI, Anthropic, Google, Meta*, Xai, Perplexity and others), collect their results and compare them honestly with each other and with the original benchmark. A fair experiment.

The results disappointed

The results fell far short of expectations. Contrary to everything, GPT 5.5 Pro not only failed to do better than the rest, it came in last. That was a shock from the start: OpenAI's flagship lost in every category of analysis, including identifying popular repositories and assessing parser performance.

Claude Opus 4.7, known for deep analysis and its ability to hold context, also failed to take first place, although its results were above average. Instead, more specialized and compact models unexpectedly came out ahead, handling the practical details of the task better.

The author admits that his manual benchmark may indeed have contained an error, in which case he chose a not-quite-optimal parser. But even so, the result shows an interesting pattern: the size and self-declared quality of an AI model do not always guarantee success on a specific practical task.

What this means

The experiment reminds developers that AI models are tools with their own strengths and weaknesses. For specific technical tasks, it is worth checking not only a model's popularity but also its actual performance on your particular case. And yes, sometimes a tool that promises to finish in 5 minutes may need your careful attention and validation.

*Meta is designated an extremist organization and banned in Russia.