[DigitalToday reporter Hwang Chi-gyu (황치규)] OpenAI has unveiled two small models designed to handle delegated tasks in AI agent systems. They are GPT-5.4 Mini and Nano.

The two models focus on codebase search, file review and parallel subtasks that need fast, low-cost processing.

The News Stack drew attention to the fact that in some areas the performance gap between the Mini model and the full GPT-5.4 model is not large. OpenAI said that in coding and computer-use benchmarks, Mini delivers performance comparable to the full model while running more than twice as fast. Nano is a simplified model specialised for large-volume work, suitable for classification, data extraction, ranking and lightweight coding support.

GPT-5.4 Mini can be used in the API, Codex and ChatGPT. Its context window is 400,000 tokens and it supports text and image inputs. GPT-5.4 Nano is available only through the API.

In SWE-bench Pro, a benchmark that evaluates real software engineering tasks, Mini scored 54.38 percent, only 3 percentage points behind the full GPT-5.4 model. In OSWorld-Verified, which measures computer-use capability, it scored 72.13 percent, close to the flagship model's 75.03 percent. Nano performs below Mini, but it outperformed the existing GPT-5 Mini in coding and tool-calling tasks.

OpenAI is highlighting an environment in Codex in which GPT-5.4 handles planning, coordination and final review, while Mini sub-agents run in parallel and take on focused tasks such as codebase search, large file review and processing related documents.

OpenAI stressed: "In such an environment, the best model is not the biggest model, but a model that responds quickly, uses tools reliably, and performs well even on complex professional tasks."