Figure 9: Share of rubric items satisfied against cost and latency per query. All points except ours are from the public FrontierFinance leaderboard. Our run used the benchmark's 220 queries and grader; its cost includes the open model, the tools and the separate model that writes the final answer.