Anthropic and get_count_tokens

[Miguel Rodríguez (miguelrgonzalez)]

I'm using Anthropic models via Bedrock and I'm getting the following warning when trying to count tokens for this model.

/app/.venv/lib/python3.12/site-packages/langchain_aws/chat_models/bedrock.py:1008: UserWarning: Falling back to default token method due to missing or incompatible `anthropic` installation (needs <=0.38.0).
2025-09-11T14:37:27.610558210Z 
2025-09-11T14:37:27.610563301Z If using `anthropic>0.38.0`, it is recommended to provide the model class with a custom_get_token_ids method implementing a more accurate tokenizer for Anthropic. For get_num_tokens, as another alternative, you can implement your own token counter method using the ChatAnthropic or AnthropicLLM classes.
2025-09-11T14:37:27.610567802Z 

I understand this is because the newer Anthropic API is not yet supported in Bedrock itself.

what's the recommended approach to get somewhat accurate token counts other than using the fallback?

[Gautam Vhavle (GautamVhavle)]

The warning is from ChatBedrock.get_token_ids, not from Bedrock itself. For Anthropic model ids it tries get_token_ids_anthropic(), which only works with anthropic<=0.38. Newer anthropic dropped that local tokenizer, so it falls back to LangChain's default (tiktoken-ish) and yells.

Accurate counts, in order of how close they are to what Bedrock will bill:

1. After the call, use the usage Bedrock already returned. AIMessage.usage_metadata / response_metadata["usage"] is the real input/output token count. If you only need it for logging or cost, do not pre-count.

2. Before the call, on current ChatBedrock, get_num_tokens_from_messages() already calls Bedrock count_tokens when count_tokens_api_supported_for_model(model_id) is true (Claude 3+ on Converse/invoke). That is the same tokenizer Bedrock uses. Prefer that over get_num_tokens(text) / get_token_ids.

3. If you are stuck on a version that still hits the warning, pass a local counter and skip the anthropic package path:

import tiktoken
enc = tiktoken.get_encoding("cl100k_base")  # close, not identical to Claude

chat = ChatBedrock(
    model_id="anthropic.claude-3-5-sonnet-20241022-v2:0",
    custom_get_token_ids=lambda text: enc.encode(text),
)

custom_get_token_ids short-circuits the anthropic<=0.38 branch. Counts will be approximate (Claude's tokenizer is not cl100k). For a tighter pre-check, call bedrock-runtime.count_tokens yourself with the same modelId you invoke.

Do not install anthropic==0.38 just to silence this. That SDK is years behind and will fight the rest of your stack. The CountTokens API / usage_metadata path is the one that matches Bedrock.