TB-MathModel: leaderboard

Agentic terminal mathematical-modeling contest tasks scored against a panel of human outstanding papers with an executable feasibility gate. TB-MathModel v1: 18 CUMCM/MCM contest problems with 21 figures run as agentic terminal tasks. O-Eval is the multi-reference score (%) against a panel of human outstanding-award papers, after a hard feasibility gate that zeroes infeasible answers; mean of three seeds per task. Model, harness and reasoning tier are bound from Appendix Table 7 (the main Table 2 omits tiers). Multi-BO is a ratio to the DeepSeek V4.1 Flash baseline and cost is USD; neither is admitted. Harness-sensitivity runs (Table 3) and the four-entry per-task Table 6 are not admitted. Question access: Source code is linked (github.com/wxj630/terminal-bench-math-modeling); public contest problems, but the graded anchor set was not verified.

Metric: Multi-reference O-Eval (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 7 models tracked.

Top models

#ModelScoreOverall rank
1GPT-6 Astra (Benchmark Sweep Run: source=TB-MathModel; harness=codex CLI; tier=max; 18 tasks; three-seed mean)64.58
2Claude Fable 5.1 (Benchmark Sweep Run: source=TB-MathModel; harness=claude-code CLI; tier=max; 18 tasks; three-seed mean)63.11
3DeepSeek V4.1 Flash (Benchmark Sweep Run: source=TB-MathModel; harness=opencode CLI; tier=high; 18 tasks; three-seed mean)59.81
4GPT-5.5 (Benchmark Sweep Run: source=TB-MathModel; harness=codex CLI; tier=xhigh; 18 tasks; three-seed mean)59.75
5Kimi K3 (Benchmark Sweep Run: source=TB-MathModel; harness=opencode CLI; tier=high; 18 tasks; three-seed mean)58.23
6GLM-5.3 (Benchmark Sweep Run: source=TB-MathModel; harness=opencode CLI; tier=high; 18 tasks; three-seed mean)57.78
7Qwen 3.8 Flash (Benchmark Sweep Run: source=TB-MathModel; harness=opencode CLI; tier=high; 18 tasks; three-seed mean)52.66

Interactive version: theaggregate.ai/benchmark?slug=tb-mathmodel · How It Works · Data refreshed daily, snapshot 2026-11-07.