Clone detectionThe models that write as one9 clusters

Cluster every pair over 90% similarity and 178 models collapse into 9 clusters. Some are siblings from the same lab. Others are strangers from different providers that converged on the exact same voice.

Cross-provider twinsDifferent labs, same writing DNA30 pairs

Strip out the same-lab siblings and 30 pairs remain that come from different providers yet write almost identically. Training-data convergence, measured, crossing every brand line.

Price arbitrageSame writing, different bill50 pairs

Pairs over 75% similarity, sorted by what the cheaper one saves you. Same voice, different bill. You are paying for the logo.

House stylesWhich labs have a recognizable voice?4 labs

Distinctiveness: how tightly a lab's own models write like each other versus like everyone else. Above 1, the lab has a fingerprint of its own. Only 4 labs clear that bar. Most providers, the largest ones included, have no detectable house style.

Family cohesionDo model families actually write alike?11 families

A family ships a flagship, a mini, a nano. Do they share a voice? The GPT-5 family is the most coherent at 71.4%. The DeepSeek line barely holds together.

What separates themThe features that tell models apart32 dimensions

Of the 32 dimensions, a few carry almost all the signal. The loudest differentiator is Sentence length variance. Structural pacing and rhythm are nearly identical across the whole catalog.

Unique vs genericWho writes unlike anything else?40 models

Average a model's similarity to every other model and you get one number: how generic it is. The unique end (lime) is mostly reasoning models writing their own dialect. The generic end clusters around the bland center of the catalog.

ConsistencyWho writes the same way every time?40 models

Style drift is the response-to-response variance inside one model. Low drift means a predictable voice no matter the prompt. High drift means the model reshapes itself per task.

Which AI models write the most similarly?

Across 178 models and 15,753 pairwise comparisons, 12 pairs write near-identically (above 90% cosine similarity on a 32-dimension fingerprint), grouping into 9 clusters. Some are same-lab siblings; some are strangers from different providers.

Is there a cheaper model that writes like Claude 3 Opus?

Gemini 2.5 Flash Lite Preview 06-17 (google) writes 78% like Claude 3 Opus (anthropic) at about 185x less per token (roughly 99% cheaper). On writing style alone the cheaper model is largely interchangeable. The price gap pays for the brand and for reasoning depth this measure does not capture.

How was AI model similarity measured?

Each model's responses to 43 standardized prompts reduce to a 32-dimension fingerprint covering lexical richness, sentence structure, punctuation, formatting, and discourse. Features are z-score normalized, then compared with cosine similarity. A pair above 0.90 is a clone (the top 0.08% of all pairs). Corpus: 3,095 responses.

Which AI provider has the most distinct writing style?

Meta has the strongest house style, writing 37.5x more like its own models than like other labs, followed by zhipu (12.3x) and deepseek (6.1x). A distinctiveness above 1 means a detectable writing signature.

Which AI model writes the most uniquely?

Gemini 2.5 Flash Preview 05-20 (thinking) (google) is the most stylistically unique model, with the lowest average similarity to every other model. Reasoning and thinking-mode models dominate the unique end; mid-tier models converge toward an average voice.

What is a stylometric fingerprint?

A stylometric fingerprint is a vector of 32 measurable writing-style features (word length, sentence variance, punctuation rates, formatting habits, hedging and transition patterns) averaged across a model's responses. It captures how a model writes, not what it says. That is why two models can share a fingerprint while differing in capability.