When they were using the Sonnet 3.5 API, they censored the word "Claude" and replaced "Anthropic" with "Meta", then later when people realized this, they removed it.
Also, after GPT-4o they switched to a llama checkpoint (probably 405B-inst), so now the tokenizer is in common (no more tokenization trick).
Also, after GPT-4o they switched to a llama checkpoint (probably 405B-inst), so now the tokenizer is in common (no more tokenization trick).