A new head-to-head comparison is pitting three large language models against one another on software-engineering tasks: GLM-5.2, DeepSeek V4 and Kimi K2.6.
According to tech-insider.org, surfaced via Google News, the three models were evaluated on their "SWE" — software-engineering — capabilities, with the write-up headlining a figure of "62% SWE Pro" for 2026.
Software-engineering benchmarks like this are meant to measure how well an AI model can handle real coding work — reading a codebase, fixing bugs and completing programming tasks — rather than just answering trivia or writing prose. A single percentage score is shorthand for how many of those coding challenges a model gets right, which is why builders and buyers increasingly treat these numbers as a rough proxy for how useful a model will be as a coding assistant.
The source frames the story as a direct comparison, though the reporting available here does not break out which of the three models posted the 62% result, nor the full methodology behind the test.
Why it matters: coding ability has become one of the most closely watched yardsticks for competing AI models, and side-by-side scorecards like this one are how the fast-moving field of rival systems gets ranked in plain numbers that non-experts can follow.