GPT-5.4 (low) is a specialized configuration of OpenAI's GPT-5.4 frontier language model, released in early 2026. This version is defined by its use of the "low" reasoning effort setting, which optimizes the model for reduced latency and lower token costs while maintaining the sophisticated architecture of the GPT-5.4 series. It is a unified model that integrates the advanced coding capabilities previously isolated in the Codex line with general-purpose reasoning, multimodal understanding, and agentic tool use.
A standout feature of the GPT-5.4 family is its native computer use capability. The model can interpret visual screenshots and issue mouse and keyboard commands to operate desktop environments and web browsers. In benchmarks such as OSWorld-Verified, the GPT-5.4 series has demonstrated the ability to perform complex, multi-step desktop tasks with a success rate surpassing human baselines. This makes the "low" configuration particularly effective for high-volume agentic workflows where responsiveness and dependable tool invocation are critical.
Technically, GPT-5.4 (low) supports a massive context window of 1,000,000 tokens, enabling it to process and reason over large-scale datasets, including entire software repositories or lengthy technical manuals. The model utilizes a chain-of-thought (CoT) process designed for high monitorability, allowing developers to inspect the model's reasoning steps for safety and alignment. Despite its high intelligence, it is tuned to be more concise and faster than the higher-effort "Thinking" or "Pro" variants of the same model family.