Stylometric analysis of 3,095 standardized responses from 178 AI models reveals nine clone clusters exceeding 90% cosine similarity on z-normalized 32-dimensional feature vectors that track lexical richness, sentence structure, punctuation, and discourse markers (rival.tips/research/model-similarity).

Primary data show Mistral Large 2 and Large 3 2512 reaching 84.8% on a composite metric of head-to-head similarity, Pearson correlations, length consistency and cross-prompt stability; Gemini 2.5 Flash Lite matches Claude 3 Opus output style at 78% while costing 185 times less. Meta exhibits the strongest provider house style with a 37.5x distinctiveness ratio. The original coverage correctly reports prompt-specific effects—satirical fake news triggers maximum convergence, letter-counting maximum divergence—yet understates the downstream implication: repeated distillation from similar base corpora is producing measurable stylistic convergence faster than architectural differentiation. Shumailov et al. (Nature, 2023) documented model collapse under recursive training on synthetic data; the observed clusters align with that trajectory, confirming homogenization patterns first quantified in that recursion study and later replicated in Stanford's 2024 stylometric survey of post-2023 LLMs.

Mainstream reporting amid the 2024-2025 hype cycle has emphasized benchmark gains and context-window size while missing the economic signal: low-cost models can now replicate premium stylistic fingerprints, eroding differentiation and complicating watermarking, detection, and creative attribution. The dataset's prompt-controlled head-to-head and per-feature Pearson signals supply primary-source evidence that current alignment techniques amplify rather than mitigate this convergence.