TB-MathModel: GPT-6 Astra tops contest bench at 64.58% O-Eval

Across 18 real problems from the CUMCM and MCM mathematical modeling contests of 2023 through 2025, GPT-6 Astra wrapped in a codex harness reached 64.58% on a multi-reference panel, with Fable 5.1 plus claude-code at 63.11%. That is the headline of TB-MathModel, a new benchmark from The Hong Kong University of Science and Technology (Guangzhou), IDEA Research and DataArcTech Ltd.

The authors describe it as, "to our knowledge the first benchmark to turn complete contest-style mathematical modeling problems into an agentic, verifiable, terminal evaluation." Each task ships the verbatim contest statement, the official attachments, and PDF figures into a Docker container capped at 2 CPUs, 8 GB of memory, no GPU, and a 4-hour timeout. The agent writes one JSON artifact and that artifact is what gets scored. Rather than promoting a single "correct" paper, the panel pulls answers from 99 Outstanding-award human papers, 94 of which supply the scored quantity, with one to 18 panel papers per task.

Cost is the other pointed finding. The two leading proprietary runs cost between $246 and $273 per 18-task pass, while a $2.54 run of DeepSeek V4.1 Flash on the opencode harness held 59.81%, which the paper reports as "1/100 the cost" of the leader for a score about five points behind. Pushing DeepSeek-Flash to its maximum reasoning tier made things worse: its multi-reference score fell to 48.11%.

Feasibility is checked twice. A programmatic gate over six decidable invariant forms covers all 18 tasks, and a Lean 4 layer re-proves the same invariants across 624 kernel-checked obligations; a failed gate zeroes a task. The gate hard-invalidated seven cells, concentrated on the heliostat problem (rated power below 60 MW, or a mirror below the 2 m minimum), plus single cells on smoke-screen, Juneau, and Olympic.

The paper is frank about its own mistakes. On the dragon-dance problem, the models converged on one value and the human Outstanding-award anchor turned out to be the physically infeasible outlier. "On dragon-dance the models were right and our anchor was wrong—the sharpest argument for independently checkable feasibility," the authors write, and that is what the Lean layer is for. Varying only the harness around DeepSeek-Flash moved O-Eval by at most 0.07 points across four harnesses, so the authors argue the board "measures the model far more than the harness it is wrapped in." TB-MathModel is the latest entry in a busy stretch of agent benchmarks we have been tracking.