AI models diverge most on coding questions, study of five models finds

A divergence study sent the same 24 questions to five AI models, including GPT, Claude, Gemini, Grok and Llama. The models agreed closely on factual questions but split sharply on coding ones, averaging 2.67 divergences per coding question. Five of the six coding questions produced at least one claim that only a single model made. Examples included Postgres partitioning, where only one model noted that retrofitting it forces a schema change, and TypeScript interfaces versus types, where lone claims included both a useful warning and a factual error. Of 14 hand-checked sole-source claims, 8 were the most valuable line in their set and 3 were wrong or invented. The author, who builds the MultiO platform, recommends checking where models disagree before committing to significant architecture decisions.