Terminal-style mathematical modeling benchmark TB-MathModel: the first agentic and verifiable mathematical modeling benchmark

Researchers released the TB-MathModel benchmark, evaluating the mathematical modeling ability of large language model agents in a containerized terminal with 18 real CUMCM/MCM competition problems; the best, GPT-6 Astra, scored only 64.6%. Agents must read the problem, view tables and images, write code, run solvers, and output structured JSON; scoring does not grade the report but only checks the numerical core of the final problem, is direction-aware, uses one-sided scoring, and does not deduct points for exceeding human papers. Each problem is paired with 94 reference answers, with each outstanding-award human paper casting one vote, and the model must be close to multiple experts simultaneously; feasibility is programmatically checked by six decidable invariants plus 624 kernel-checked Lean 4 proof obligations, without relying on Mathlib and without leaving any sorry. Claude Fable 5.1 scored 63.1%, and an open-source model costing 2.54 USD achieved 59.8% at 1/100 the cost.