OpenAI has just launched a surprise release of two lightweight killers in the GPT-5.4 series: GPT-5.4 mini and GPT-5.4 nano! Performance skyrockets, costs plummet. If you are working on code development or multi-agent systems, this is definitely the most important infrastructure upgrade of the year. GPT-5.4 mini: The strongest "small core" and perfect candidate for high-order sub agents. This is OpenAI's strongest small model to date, with a 400K context window. Compared to the previous generation 5 mini, the speed has been directly increased by more than 2 times! In benchmark tests such as SWE bench Pro, the performance has approached the full blooded version 5.4. Core killer move: one ️⃣ Extremely low latency Vibe Coding experience: In the high-frequency iterative code workflow (directed editing, code library navigation, front-end generation, debugging), the response is extremely fast. two ️⃣ Multimodal computer control: capable of quickly parsing complex GUI (graphical user interface) screenshots and performing computer vision operations. three ️⃣ The perfect multi-agent puzzle: the best practice provided by the official - let GPT-5.4 be the "main brain" responsible for high-level planning, distributing high concurrency tasks such as searching the code base and reviewing large files to the 5.4 mini as a "sub agent" to run, maximizing efficiency! API pricing: Input $0.75/1M tokens and output $4.50/1M tokens. (Called in Codex, only 5.4% of the quota is consumed) GPT-5.4 nano: pushing "low cost and high speed" to the extreme Specially designed for tasks with "abnormal" requirements for latency and cost. A significant upgrade has been made to GPT-5 nano, which is currently only available in the API. Recommended scenario: A cleaner for massive basic tasks. Extremely suitable for data classification, information extraction, sorting, and handling the most basic auxiliary tasks in complex systems. API pricing (outrageously cheap): Input $0.20/1M tokens and output $1.25/1M tokens. To summarize in one sentence Choose 5.4 mini: Pursuing extremely low latency, but at the same time requiring strong reasoning, coding ability, and multimodal understanding. Select 5.4 nano: Processing massive, simple, and repetitive basic data cleaning and auxiliary tasks. 5.4 Mini has been fully launched in API, Codex, and ChatGPT (free version and Go users can find "Thinking" in the "+" menu). 5.4 Nano is API only.