OpenAI rolls out smaller models to step up its push for intelligent agents, quietly launching a “price war” with domestic models

In the age of agents, the industry is no longer chasing flagship models with ever-larger parameters; instead, it is rolling out more, faster, and more efficient small models. On March 18 Beijing time, OpenAI announced the release of two small models, GPT-5.4 mini and nano. The company says these are its internally “most capable small models to date.” Their capabilities are close to the flagship model GPT-5.4, but they have been optimized for high-frequency workloads, aiming to provide new options for agent-era applications with lower latency and higher cost-performance. Industry analysis believes this is a key step for OpenAI to fill in the missing pieces of its product puzzle for the agent era. After AI moves into real business, not every step needs to “use a sledgehammer to crack a nut.” These small models are precisely intended for the execution layer, serving as the main power for sub-agents. However, when it comes to pure cost-performance, this battlefield has long been shrouded in gun smoke—especially with China’s leading models holding a dominant position. Overseas, some developers have said that GPT-5.4 mini “failed upon release,” because China’s Kimi-K2.5 model is not only priced lower, but also performs better. Still, others believe the reference value of benchmark tests is declining, and the real winner must be verified in real-world tasks. (First Financial)