Anthropic says Opus 4.7 uses "an updated tokenizer that improves how the model processes text", but "the tradeoff is that the same input can map to more tokens"