DeepSeek V4 Pro Matches GPT‑5.2 on FoodTruck Bench, 17× Cheaper

DeepSeek V4 Pro, the first Chinese model to reach frontier tier on the FoodTruck Bench, matched GPT‑5.2’s median outcome within 3 % after a 10‑week gap since GPT‑5.2’s mid‑February test. The benchmark, which simulates a food‑truck operation with 34 tools and persistent memory, places DeepSeek at #4 overall behind Opus 4.6, GPT‑5.2, and Grok 4.3.

Pricing is a major differentiator: GPT‑5.2 charges $1.75/M input and $14/M output, while DeepSeek runs at $0.435/M input and $0.87/M output, making it roughly 17× cheaper for the same agentic workload. In cost‑efficiency terms, DeepSeek ranks #2 on the leaderboard, only behind Gemma 4 31B, and outperforms Grok 4.3 in consistency—zero losses, six times less food waste, 30 % more meals per day, and a 2.4× tighter outcome distribution. Opus 4.6 still achieves a higher peak, and Gemma remains the cheapest option. Xiaomi MiMo v2.5 Pro also entered the top six, posting a 1,019 % median ROI and $22,388 median net worth at $2.41/run, though its variance is wider. Together, these results show two Chinese models in the top six, both under $3.5/run, a sharp shift from the frontier gap of a year to just ten weeks.