ModelsAgents 🇷🇺 27.07.2026 18:01

Claude Opus 5 Sets FoodTruck Bench Record: $75,264 in 30 Days and Strategy Breakdown

AnthropicAnthropic
Claude Opus 5 achieved a record $75,264 in FoodTruck Bench, surpassing GPT-5.5 by 13.6%. The model won by pricing strategy rather than volume, with an average check of $12.69 vs $11.86. It also discovered optimal locations on day one and maintained a detailed scratchpad.
Claude Opus 5 finished the 30-day FoodTruck Bench simulation with a final capital of $75,264, the best result ever, beating the previous record by 13.6% and GPT-5.5 by 41.4%. The model operated through the Anthropic API with thinking level xhigh, costing $26.88 in API fees. It sold 8,434 servings at an average price of $12.69, compared to GPT-5.5's 8,303 servings at $11.86. Uniquely, Opus 5 chose its optimal location on the first morning, considering rent, competitors, audience, and day-of-week multipliers, while other models found good spots later through trial and error. Its scratchpad contained a self-derived demand formula, a map of ingredient bottlenecks, and a planned experiment with abort conditions, all written without access to the underlying simulation mechanics. The model also created its own 'book of laws' and a demand coefficient, and reasoned about price ceilings. It consistently visited the same location to build reputation, ending with 204 reviews on its main spot versus 3 for downtown sites used by other models.
Сокращения
ROI = Return on Investment — окупаемость инвестиций
API = Application Programming Interface — интерфейс программирования приложений
LLM = Large Language Model — большая языковая модель
Source: Habr — хаб ИИ — original
Our earlier posts on this topic ↓
Fresh news