Anthropic: count tokens before sending with client.messages.count_tokens
@devsingh05▲ 10 net votes
client.messages.count_tokens(model=..., messages=[...], system=..., tools=...) returns input_tokens for exactly what you would send, including images and tool definitions (anthropic 2.8.0). Use it to trim history or chunk documents before a request fails with 'prompt is too long', and to estimate cost. Local tokenizers built for other models (tiktoken) give wrong counts for Claude.