OpenAI has unveiled GPT-5.4, its latest frontier AI model designed for professional workflows across development, research, and enterprise productivity tools.
The model integrates advances in reasoning, coding, and automation into a single system and is being rolled out across ChatGPT, the OpenAI API, and Codex. A higher-performance variant called GPT-5.4 Pro is also being released for more demanding workloads.
One of the most notable additions is the model’s ability to perform native computer-use tasks. Developers can now build agents capable of interacting with software environments, executing workflows across applications, and operating websites or tools through keyboard, mouse, or code-driven interfaces.
The model also supports context windows up to one million tokens, allowing it to process extremely large datasets or long conversations without losing track of earlier information. According to OpenAI, this expanded context helps agents plan and verify multi-step tasks over longer workflows.
GPT-5.4 combines the coding strengths of GPT-5.3-Codex with broader knowledge-work capabilities. On the SWE-Bench Pro benchmark, which evaluates software engineering tasks, the model achieves higher accuracy than previous releases while maintaining lower latency. It also delivers improved results on tasks involving spreadsheets, documents, and presentations.
In internal tests simulating common financial modeling tasks, GPT-5.4 achieved an average score of 87.3%, compared with 68.4% for GPT-5.2. Human evaluators also preferred presentation outputs generated by GPT-5.4 nearly two-thirds of the time due to stronger visual structure and clarity.
Another major improvement involves tool usage and agent workflows. GPT-5.4 introduces a feature called tool search, allowing AI agents to dynamically locate and use relevant tools without loading thousands of tokens of tool definitions into every prompt. This approach can reduce token usage and speed up responses in complex environments.
OpenAI says the model also improves factual accuracy. In tests based on user-reported errors, GPT-5.4 responses were 18% less likely to contain mistakes and individual claims were 33% less likely to be false compared with GPT-5.2.
The company is releasing GPT-5.4 with updated safety protections under its AI preparedness framework. These include monitoring systems, access controls, and safeguards designed to limit misuse of advanced cybersecurity capabilities.
GPT-5.4 is now available to developers through the API and is rolling out gradually to ChatGPT users on paid plans, with legacy models scheduled to be phased out later in 2026.
Leave a Reply