Research Figures
Eight figures. Full empirical record.
Figure 1
Token Counts by File Type and Size
Token counts across Small, Medium, and Large files for all five types. PDFs and DOCX files consistently carry the heaviest token loads — a large PDF reaches 38.5K tokens per turn alone.
Figure 2
The Compounding Effect Across Turns
Cumulative token growth for Small, Medium, and Large PDFs across 44 turns. The gap between file tokens (orange) and message tokens (dashed) widens relentlessly with every turn.
Figure 3
Rₑ Heatmap: File Token Proportion
Proportion of session input tokens attributable to the uploaded file, across all file types and conversation lengths. Near-uniform deep red confirms file dominance exceeds 99% in every scenario tested.
Figure 4
Token Composition: Stacked Area View
Stacked area chart for a medium PDF over 44 turns. File tokens (orange) occupy virtually the entire stack — user messages and system prompts form an invisible sliver at the bottom.
Figure 5
Session Cost Breakdown in USD
Left: dollar cost per component in a 20-turn session. Right: pie chart of input token cost distribution. File tokens account for over 99% of all input spend.
Figure 6
Rₑ by File Size — All Types
File token proportion as a function of file size for all five types at 10 turns. All lines cluster tightly above 99%, confirming that even small files dominate once attached.
Figure 7
Full Cost Scaling Grid
Session input cost (USD) heatmap across all file types, sizes, and turn counts up to 80 turns. Large files in long sessions reveal how costs compound rapidly in production.
Figure 8
Prompt Caching: Cost Savings
Top: cumulative cost with vs without prompt caching. Bottom: percentage savings curve. By turn 20, caching delivers 84.2% cost reduction — the single highest-impact optimization available.