I am running the same instructions every time but on different user input, so I am using the beta cache feature to reduce the input token amount

No matter how I structure my prompts It will not properly cache, I have only been able to go from 1700 input tokens to 1600.

My goal is to keep the instructions and more_info in the cache.

Parts of the code are filled with random bits for privacy.

from config import ANTHROPIC_API_KEY
import json
import time
import anthropic
import pandas as pd

client = anthropic.Client(api_key=ANTHROPIC_API_KEY)

# Load the full dataset
input_file = "data.csv"
output_file = "data.json"

df = pd.read_csv(input_file, header=None)
df.columns = ["1", "2"]

# Define the system prompt to set Claude's role
system_prompt = """
insert here
"""

# Define the constant parts of the user prompt
instructions = """
long instructions
"""

more_info = """
a lot of info
"""

# Define the output format and prefill Claude's response
PREFILL = """
{
    prefill
}
"""


def get_completion(task: str, system_prompt="", prefill=None):
    response = client.beta.prompt_caching.messages.create(
        model="claude-3-5-sonnet-20240620",
        system=system_prompt.strip(),
        max_tokens=1000,
        messages=[
            {
                "role": "user",
                "content": [
                    {
                        "type": "text",
                        "text": f"{instructions}nn{more_info}",
                        "cache_control": {"type": "ephemeral"},
                    },
                    {
                        "type": "text",
                        "text": f"<task>{task}</task>",
                    },
                    {
                        "type": "text",
                        "text": "short sentence",
                    },
                ],
            }
        ],
        temperature=0.0,
    )
    return response.content[0].text.strip()


def main():
    try:
        start_time_total = time.time()
        tasks_output = []

        for idx, row in df.iterrows():
            task_id = row["Task ID"]
            task_description = row["Task Description"]

            if task_description:
                start_time = time.time()

                assistant_reply = get_completion(
                    task=task,
                    system_prompt=system_prompt,
                    prefill=PREFILL,
                )

                end_time = time.time()

                # Try to parse the assistant's reply as JSON
                try:
                    response_json = json.loads(assistant_reply)

                    # Add the task ID to the JSON output
                    task_json = {
                        "Task Number": task_id,
                        **response_json,
                    }

                    tasks_output.append(task_json)

                    print(
                        f"Processed task {task_id}: {task_description} (Time taken: {end_time - start_time:.2f} seconds)"
                    )

                except json.JSONDecodeError:
                    print(
                        f"Error: Invalid JSON response for task {task_id}. Moving on..."
                    )
                    print(f"Raw Model Output: {assistant_reply}")

                    task_json = {
                        "Task Number": task_id,
                        "Invalid JSON Response": assistant_reply,
                    }
                    tasks_output.append(task_json)

        # Save the results to the output JSON file
        with open(output_file, "w") as outfile:
            json.dump(tasks_output, outfile, indent=4)

        end_time_total = time.time()
        time_taken_total = end_time_total - start_time_total
        minutes, seconds = divmod(time_taken_total, 60)
        print(
            f"All tasks processed. Results saved to {output_file} nTime taken: {int(minutes)} minutes and {seconds:.2f} seconds"
        )

    except Exception as e:
        print(f"Error: {e}")


if __name__ == "__main__":
    main()

I have tried structuring the prompts in various ways, combining all into one static prompt and caching it. No luck.