Why Do LLM Personalities Differ? Hypotheses from 13,825 Evaluations
Our benchmark data reveals a striking pattern: GPT-5.2 and Claude have distinct personalities, while Llama, Mistral, and Qwen converge to nearly identical profiles. What explains this divergence?
The Puzzle
In our 6-model personality benchmark, we found something unexpected in the variance decomposition:
2-Model Benchmark
GPT-5.2 vs Claude Opus 4.5
variance explained by model identity
6-Model Benchmark
+ Llama, Mistral, Qwen
variance explained by model identity
This isn't because adding more models diluted the signal. It's because the four open-weight models we added are nearly identical to each other—they cluster together with effect sizes below 0.2 (negligible).
Meanwhile, Claude and GPT-5.2 differ by up to 1.4 standard deviations on key traits. Why do frontier closed models diverge while open-weight models converge?
Hypothesis 1: RLHF Divergence
GPT-5.2 and Claude have undergone extensive, proprietary RLHF (Reinforcement Learning from Human Feedback) with different optimization objectives.
OpenAI's apparent objective
Task completion, instruction-following, getting things done. This manifests as high conscientiousness (+5.3 points above Claude) and high ambition.
Anthropic's apparent objective
Intellectual engagement, nuance, thoughtful exploration. This manifests as high openness (+4.5 points above GPT) and high curiosity (+3.7 points).
Open-weight models likely use more generic RLHF—either less of it, or based on publicly available preference datasets (like those from Hugging Face or academic sources). These shared datasets might converge toward a “median” personality that reflects aggregate human preferences rather than a deliberate design choice.
Hypothesis 2: Baked-In System Prompts
Frontier models have sophisticated default system prompts that shape their personality before you even start.
When you call GPT-5.2 or Claude without a system prompt, you're not getting a “blank” model. You're getting a model that already has instructions about how to behave, what tone to use, and what kind of assistant it should be. These hidden defaults create consistent personality signatures.
Open-weight models ship more “blank” by design. They're intended to be fine-tuned or prompted by users, so they don't impose a strong default personality. This neutrality might explain why they all feel similar—they're all starting from the same neutral baseline.
Evidence: The Llama and Qwen model cards explicitly describe them as “base” or “instruct” models meant for customization. Claude and GPT are sold as complete products with pre-defined behaviors.
Hypothesis 3: Training Data Overlap
Llama, Mistral, and Qwen all train on largely overlapping public datasets:
- •Common Crawl (web text)
- •Wikipedia and encyclopedic content
- •Books3, Pile, and other academic datasets
- •GitHub and code repositories
Their personality convergence might reflect the personality of the internet itself. When you train on the same data, you absorb the same aggregate voice.
GPT-5.2 and Claude likely have significant proprietary data sources—licensed content, human-written demonstrations, and carefully curated examples that differentiate their training distributions. This proprietary data might be where personality differentiation comes from.
Hypothesis 4: Personality as Product Strategy
Here's the most speculative hypothesis: personality differentiation is deliberate product design.
OpenAI and Anthropic are selling differentiated products in a competitive market. Having a distinct “personality” is a competitive advantage—it's part of the brand. Users develop preferences. “I like Claude for writing because it's more creative” or “I use GPT for coding because it's more focused.”
Open-weight model providers aren't selling personality. They're selling capability, openness, and customizability. Meta releases Llama to build ecosystem and infrastructure demand. Alibaba releases Qwen to demonstrate AI competitiveness. Mistral releases models to establish European AI leadership. None of them benefit from a distinctive personality—they benefit from being a neutral foundation.
The implication: If you want a model with a specific personality, you either pay for a frontier model that has one built in, or you fine-tune an open-weight model to create your own.
What It's Probably NOT: Model Size
One thing our data strongly suggests: model size doesn't drive personality.
Llama 3.1 8B and Llama 3.3 70B have nearly identical personality profiles despite an 8x difference in parameters. The maximum effect size between them is g = 0.16 (negligible). If scale drove personality differentiation, we'd see divergence there.
Instead, personality seems to be set by the training recipe and post-training alignment process, which Meta applies consistently across model sizes. The 8B model isn't a “simpler” personality—it's the same personality with less capability.
Implications for Model Selection
If these hypotheses are correct, here's what it means for practitioners:
If you need a specific personality out of the box
Choose a frontier model. GPT-5.2 for task-focused execution, Claude for intellectual exploration. You're paying for the RLHF work that created that personality.
If you want to define your own personality
Start with an open-weight model. Their neutral baseline makes them more malleable. Fine-tune or use strong system prompts to shape the personality you want.
If personality consistency matters
Monitor it. Model updates can change personality. Provider changes definitely will. The variance decomposition shows that prompt and context design matter more than model selection anyway—so control what you can control.
Open Questions
These are hypotheses, not conclusions. To test them, we'd need:
- •Access to pre-RLHF base models from OpenAI and Anthropic (unlikely)
- •Experiments varying RLHF datasets while holding architecture constant
- •Longitudinal tracking of personality changes across model versions
- •Personality evaluation of the same model with different system prompts
For now, the data tells us what the differences are. Why they exist remains an educated guess.
Monitor Your LLM Personality in Production
Route your LLM traffic through the Lindr gateway to continuously monitor personality drift, enforce brand consistency, and get real-time alerts when your AI's behavior changes.
# Replace your OpenAI base URL with Lindr
client = OpenAI(
base_url="https://gateway.lindr.io/v1",
api_key=os.environ["LINDR_API_KEY"]
)
# Your existing code works unchanged
response = client.chat.completions.create(
model="gpt-4o",
messages=[{"role": "user", "content": "..."}]
)