Anthropic Claude prompt compression

Claude cost context

Claude 3.5 Sonnet costs $3.00 per million input tokens, and Claude 3.5 Opus costs $15.00 per million input tokens. A RAG application processing 10,000 tokens per query with 10,000 daily queries costs $300/day on Sonnet or $1,500/day on Opus.

SuperCompress reduces input tokens by ~65%, dropping those costs to ~$105/day and ~$525/day respectively. For teams running Claude at scale, this is a direct bottom-line improvement.

Anthropic wrapper integration

from anthropic import Anthropic
from supercompress import Compressor

class SuperCompressAnthropic(Anthropic):
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
        self._compressor = Compressor()

    def messages_create(self, *, model, messages, **kwargs):
        if len(messages) > 1:
            # Compress the conversation history
            history_lines = []
            for msg in messages[:-1]:
                if isinstance(msg.get("content"), str):
                    history_lines.append(msg["content"])
            history = "
".join(history_lines)
            query = messages[-1].get("content", "") if isinstance(messages[-1].get("content"), str) else ""
            if history and query:
                result = self._compressor.compress(history, query)
                # Replace history with compressed single message
                compressed_messages = [{"role": "user", "content": result.compressed_text}]
                compressed_messages.append(messages[-1])
                kwargs["messages"] = compressed_messages
        return super().messages_create(model=model, messages=kwargs.get("messages", messages), **{k:v for k,v in kwargs.items() if k != "messages"})

Frequently asked questions

Does compression work with Claude's extended thinking?

Yes. Compression happens before the API call. The thinking budget is unaffected.

Will it work with Claude 3 Haiku for cheaper calls?

Yes. Compression benefits all Claude models equally.