modern LLMs all resemble either GPT and Claude in some way, cheaper alternatives accelerate adoption

A detailed technical experiment has gone viral, revealing a surprising convergence in the output style of today's leading large language models. Researchers fed identical prompts to 25 different LLMs—including GPT-5.x, Claude Opus/Sonnet/Haiku, Grok 4.x, Gemini 3.x, DeepSeek, and Qwen—and used Google's Gemma 4 model to analyze their internal 'thought vectors.' By extracting and averaging residual stream activations across all 42 layers of Gemma 4, they created a massive 107,520-dimensional 'style vector' for each model's responses. A cosine similarity heatmap of these vectors showed a stark, two-cluster split, with one red/orange block representing a 'GPT resemblance' family and another representing a 'Claude resemblance' family.

This finding suggests that despite different architectures and companies, LLM outputs are becoming stylistically homogenized. Notable exceptions included Claude Haiku 4.5, which sat in the middle, and Gemini 3 Flash, which was a clear outlier. The pattern held even when tested with real user prompts. The implication is profound: for many general-use cases, the capabilities of expensive frontier models and cheaper alternatives are becoming difficult to distinguish. This accelerates the adoption of cost-effective, locally runnable, or open-source models, pushing the industry toward a commoditized landscape where brand and price, not just raw capability, drive user choice.