TPS-CalcBench: leaderboard
Metric: Composite KPI (0-100): mean rubric-weighted process score (formula selection, parameters, units, calculation accuracy, plausibility, assumptions, interpretation, presentation, each judged 0 to 2 by a calibrated Gemini-3-Pro-Preview judge against the reference solution) over the successfully judged items, on the 420-item high-confidence core set of TPS-CalcBench (analytical calculation problems in hypersonic thermal protection system engineering), zero-shot text-only prompting; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 12 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | GPT-5.2 | 87.85 |
| 2 | GPT-5.1 | 82.18 |
| 3 | GLM-5 | 79.82 |
| 4 | DeepSeek V3.1 | 79.22 |
| 5 | Claude Opus 4.5 | 77.66 |
| 6 | Seed 1.8 | 75.69 |
| 7 | GLM-4.7 | 75.21 |
| 8 | Claude Sonnet 4.5 | 71.75 |
| 9 | Claude Haiku 4.5 | 62.76 |
| 10 | Gemini 3 Flash (Preview) | 42 |
| 11 | MiniMax-M2.5 | 12.98 |
| 12 | MiniMax-M2.1 | 12.85 |
Interactive version: theaggregate.ai/benchmark?slug=tps-calcbench · How It Works · Data refreshed daily, snapshot 2026-10-07.