GPT-5.4 is a multimodal large language model in the GPT family, released on March 5, 2026, and distributed through NanoGPT under the openai/gpt-5.4 routing key. Third-party reporting and an Ars Technica news thread both attribute the model to OpenAI, framing it as a significant update to the GPT line focused on expanding knowledge work capability. The release introduced three tiers—Standard, Thinking, and Pro—designed to balance cost and capability across everyday production tasks, multi-step reasoning, and maximum-performance professional workloads.

Beyond the variant structure, GPT-5.4 is positioned around three practical advances. It reports a 75% score on the OSWorld computer-use benchmark, surpassing the 72.4% human baseline, and it can browse the web, interact with desktop applications, fill forms, and execute multi-step workflows autonomously without external frameworks. A dynamic tool search mechanism discovers and loads only relevant tools from a pool of more than 50 options, reducing prompt bloat by 47% and improving tool selection accuracy. The model also claims 33% fewer factual errors compared to its predecessor, and pairs a 272K native API context with Codex-driven extensions toward a one-million-token window for processing whole codebases and large document collections. These traits make GPT-5.4 well suited to complex automation, agentic tool use, and long-context analytical tasks.