Andon Labs' latest Vending-Bench 2 benchmark puts open-source model GLM 5.2 in second place. The test simulates a year-long vending machine business, feeding the model daily inventory and financial data, letting it make restocking and pricing decisions. The goal: measure decision-making consistency across long time horizons. Data shows GLM's performance has improved linearly over versions, with average monthly profit gains of nearly $1,000 (GLM 5 scored $4,432, GLM 5.1 climbed to $5,634).
https://twitter.com/andonlabs/status/2070203811776614590
Meanwhile, other major Chinese models showed diverging results. Kimi K2.7 Code slipped compared to its predecessor Kimi K2.6. Minimax M3 made a notable leap over M2.5, but still lags far behind the Kimi and GLM series in overall profitability.