DeepSeek Achieves GPT-5.2-Level Performance

In testing by FoodTruck Bench, an agentic performance measurement tool, DeepSeek V4 Pro entered frontier-model territory, posting results on par with GPT-5.2.

Key findings from the analysis are as follows:

• Performance and Ranking: It ranked 4th overall. While its results were similar to Grok 4.3, it held an edge in execution consistency, such as reducing food waste and increasing the number of meals served.
• Overwhelming Cost Efficiency: It delivers equivalent performance at roughly 17 times lower cost than GPT-5.2 on agentic workloads. In terms of cost efficiency (net worth relative to API spend), it ranked 2nd overall, behind Gemma 4 31B.
• Narrowing Technology Gap: The gap between US and Chinese models, once around a year, has now narrowed to roughly 10 weeks.
• Chinese Models on the Rise: Xiaomi MiMo v2.5 Pro also placed 6th on the benchmark, showing that lower-priced Chinese models are proving highly competitive in the agentic market.