Roughly 994 tokens appear almost always dedicated across AI labs’ BPE-like tokenizers

A user comparing tokenizers from Google, DeepSeek, Meta, Anthropic and others says about 994 tokens are almost always dedicated in BPE-like tokenizers, including across Anthropic model generations. Separately, another user says GPT-2’s regex has explicit rules for common English contractions, while GPT-4 uses case-insensitive contraction rules and tokenizes digits in groups of three.