CSV tokenization sizes are calculated using Claude's standard byte-pair tokenizer with the following formula: Tokens = Delimiters + Cell Values + System Prompt

• Delimiter Overhead: Commas and newlines consume 1 token each. For a spreadsheet, this equals (rows × columns) + rows tokens.
• Numerical Tokenization: Numbers are less well tokenized than natural language. Multi-digit numbers and floating-point values (e.g. 123.456) are split, averaging 2.8 tokens per cell value.
• Instructions & Context: Adds an extra 1,500 - 2,000 tokens for system prompt wrapper instructions, schema definition headers, and formatting rules.
• Prompt Caching Limits: Caching has a minimum threshold of 20,000 tokens. Spreadsheets below this scale (e.g. Small) pay the full base rate, whereas cached datasets pay the Cache Hit price (90% discount).
• Context Exhaustion: Large spreadsheets exceed Claude's 200,000 token context window. Directly uploading them is impossible. Carla HQ compiles models offline via WebGPU, making sheet sizes virtually unlimited.

While Claude's Prompt Caching offers a 90% discount on cache hits, real-world spreadsheet workflows rarely experience high hit rates due to several key constraints:

• ⏳ 5-Minute Expiration: The prompt cache has a short Time-To-Live (5 minutes). Typical analyst queries are spaced out by several minutes or hours, leading to frequent cache expirations and constant write cycles.
• 🔄 Cell Edits Invalidate the Cache: Caching matches the exact prefix hash of the prompt. If you edit a single cell, insert a row, or filter/sort columns, the entire CSV string changes, resulting in a 100% cache miss.
• 💳 Cache Write Surcharge: Writing a new cache costs 25% more than the standard input rate ($3.75 / MTok vs $3.00 / MTok on Sonnet 5). If your cache is frequently invalidated, your total API cost will exceed base rates.
• 🚀 Carla HQ Bypasses Caching: Carla HQ compiles model weights locally using WebGPU. Cell edits update the local model instantly. The MCP connection only sends lightweight prediction requests (~250 tokens), avoiding caching constraints entirely.

An interaction represents a single query, forecast, or prediction request processed by the LLM. In typical sheet workloads, these interactions correspond to:

• 💬 Conversational Chats: Asking the sidebar assistant to summarize a selected table, highlight key trends, or explain a cell's prediction formula.
• 🏷️ Row-by-Row Labeling (Macros): Dragging down a prediction formula across a column (e.g. sentiment classification, categorizing customer feedback, normalizing addresses).
• 📊 Bulk Row Forecasting: Predicting missing values in a data grid based on columns that have been mapped to local WebGPU trained regression parameters.

Typical Usage Scenarios:

• 🧑‍💻 Individual Analyst Work: A single analyst exploring cells, generating predictions, and asking chat questions averages 50–100 interactions per day (approx. 2 weeks of usage for 1,000 interactions).
• ⚡ Automated Batch Loop: Classifying a standard 1,000-row survey feedback sheet executes exactly 1,000 interactions back-to-back. Under Naive CSV prompting, this costs $75.00 and exhausts API limits, while Carla HQ costs only $0.75.
• 👥 Team Collaboration: A small team of 5 analysts actively modeling sales data will easily run 1,000 interactions per day.