OpenAI’s GPT-5.4 doubles down on safety as competition heats up
In the midst of recent developments and controversies surrounding a contract with the U.S. Department of Defense, OpenAI released the GPT-5.4 model.
The release comes at a time when users are reportedly leaving ChatGPT for rival chatbots, particularly Anthropic’s Claude.
GPT-5.4 is rolling out gradually across ChatGPT and Codex and is available through the API as gpt-5.4. In ChatGPT, GPT-5.4 Thinking is available to Plus, Team, and Pro users, while Enterprise and Edu customers can enable early access. GPT-5.4 Pro is available through the API as gpt-5.4-pro and to Pro and Enterprise plans.
In its announcement, OpenAI described GPT-5.4 as a flagship model that combines recent advances in reasoning, coding, and agent-based workflows. The company said the model expands on GPT-5.3-Codex and improves how it handles tools, software environments, and everyday office tasks.
They claim that the new model improves performance on professional tasks including spreadsheets, presentations, and document handling, and adds stronger coding capabilities, support for more complex multi-step tasks, a larger context window, and improved factual accuracy with fewer hallucinations.
“To make GPT-5.4 better at real-world work, we continued our progress in reducing hallucinations and errors. GPT-5.4 is our most factual model yet: on a set of de-identified prompts where users flagged factual errors, GPT-5.4’s individual claims are 33% less likely to be false and its full responses are 18% less likely to contain any errors, relative to GPT-5.2,” OpenAI stated.
OpenAI also said GPT-5.4 is its first general-purpose model with native computer-use capabilities and described it as a major step forward for developers and AI agents.
Safety protections upgraded
OpenAI reported it strengthened safeguards while preparing GPT-5.4 for release, keeping the same high cyber-risk classification used for GPT-5.3-Codex and deploying additional protections outlined in its system documentation. These include expanded cyber safety systems, monitoring tools, trusted access controls, and request blocking for higher-risk activity on Zero Data Retention surfaces.
The company described its approach as precautionary due to the dual-use nature of cybersecurity capabilities and acknowledged that detection systems may still produce false positives. The updates aim to reduce unnecessary refusals and maintain protections against misuse.
Separately, OpenAI published new safety research on monitoring how models reason, including an open-source evaluation designed to test whether systems can conceal their reasoning. The research found that GPT-5.4 Thinking showed a low ability to obscure its reasoning, which the company characterized as a positive safety signal.