% Generated by paper_plot_script/make_figures_tex.py -- do not edit.
% The caption text and the table numbers live in that chart's plot script;
% re-run the script to update this file.
\begin{table}[!t]
\centering
\small
\caption{\textbf{What one pass cost each arm.} List rate $\times$ the tokens it consumed. The two $p$ columns test each manager arm against its own single call and against Fable 5's single call: Welch over the five passes, Holm-corrected across 11 comparisons. Qwen3.8-27B's single arm is the 128k cap-matched one; at 250k it costs \$26.95/pass. Retried and discarded attempts are counted: they were generated and would be billed. The cheapest arm is GPT-5.6-Luna single at \$0.41 a pass and the most accurate is Qwen3.8-Flash-Next manager at \$5.76 --- a 14$\times$ spread in price for +22.6 points.}
\label{tab:cost}
\begin{tabular}{@{}l@{\hspace{0.5em}}l@{\hspace{0.5em}}r@{\hspace{0.5em}}r@{\hspace{0.5em}}r@{\hspace{0.5em}}r@{\hspace{0.5em}}c@{\hspace{0.5em}}c@{}}
\toprule
\textbf{Arm} & \textbf{\shortstack{Rate \$/MTok\\in / out}} & \textbf{In (MTok)} & \textbf{Out (MTok)} & \textbf{\$/pass} & \textbf{\$/solved} & \textbf{\shortstack{$p$ vs\\single}} & \textbf{\shortstack{$p$ vs\\Fable 5}} \\
\midrule
\shortstack[l]{Qwen3.8-27B\\single} & \$0.21 / \$2.55 & 0.0753 & 7.4247 & \$18.95 & \$0.28 & --- & --- \\
\shortstack[l]{Qwen3.8-27B\\manager} & \$0.21 / \$2.55 & 1.5053 & 18.6277 & \$47.82 & \$0.52 & $1.5\times10^{-4}$ & $6.1\times10^{-4}$ \\
\shortstack[l]{Qwen3.8-Flash-Next\\single} & \$0.15 / \$0.47 & 0.0753 & 5.7751 & \$2.73 & \$0.032 & --- & --- \\
\shortstack[l]{Qwen3.8-Flash-Next\\manager} & \$0.15 / \$0.47 & 1.2930 & 11.8380 & \$5.76 & \$0.062 & $3.5\times10^{-4}$ & $3.6\times10^{-7}$ \\
\shortstack[l]{DeepSeek-V4.1-Flash\\single} & \$0.30 / \$1.20 & 0.0680 & 2.8626 & \$3.46 & \$0.039 & --- & --- \\
\shortstack[l]{DeepSeek-V4.1-Flash\\manager} & \$0.30 / \$1.20 & 1.4993 & 12.4873 & \$15.43 & \$0.17 & $1.6\times10^{-4}$ & $2.2\times10^{-9}$ \\
\shortstack[l]{GPT-5.6-Luna\\single} & \$0.20 / \$1.20 & 0.0661 & 0.3299 & \$0.41 & \$0.006 & --- & --- \\
\shortstack[l]{GPT-5.6-Luna\\manager} & \$0.20 / \$1.20 & 1.1686 & 1.0522 & \$1.50 & \$0.018 & $6.4\times10^{-5}$ & $1.3\times10^{-6}$ \\
\shortstack[l]{GPT-5.6-Terra\\single} & \$2 / \$12 & 0.0661 & 0.2728 & \$3.41 & \$0.042 & --- & --- \\
\shortstack[l]{GPT-5.6-Terra\\manager} & \$2 / \$12 & 1.1098 & 0.7911 & \$11.71 & \$0.13 & $1.6\times10^{-4}$ & $1.9\times10^{-8}$ \\
\shortstack[l]{Fable 5\\single} & \$10 / \$50 & 0.0899 & 1.2043 & \$61.11 & \$0.68 & --- & reference \\
\bottomrule
\end{tabular}
\par\vspace{0.6ex}
{\footnotesize List rates: Qwen3.8-27B \citep{openrouter2026}, Qwen3.8-Flash-Next \citep{openrouter2026flash}, GPT-5.6-Luna and GPT-5.6-Terra \citep{openai2026price}, DeepSeek-V4.1-Flash \citep{deepseek2026price}, Fable~5 \citep{anthropic2026price}. OpenAI uses short-context rates and Anthropic and DeepSeek use the base rates; no long-context tier, batch/off-peak discount, or prompt-caching multiplier is applied. Qwen and DeepSeek rates were read on 2026-09-15.}
\end{table}