OpenAI unveils GPT-5.4: A new era of professional AI assistants

OpenAI officially announced the launch of GPT-5.4, positioning it as their most capable and highly efficient frontier model explicitly designed for professional environments. Available immediately across ChatGPT, the API, and Codex, this new iteration represents a substantial leap forward in the evolution of artificial intelligence. Accompanying the standard release is GPT-5.4 Pro, a premium tier aimed at users who require maximum computational power for complex tasks. This launch signifies a transition from traditional conversational artificial intelligence to comprehensive digital workers capable of managing intricate professional workflows with unprecedented autonomy.

The core architecture of GPT-5.4 brings together the absolute best of recent technological advancements in general reasoning, coding proficiency, and agentic behavior into a single system. OpenAI has successfully integrated the industry leading coding prowess of previous models while fundamentally improving how the intelligence interacts with everyday office ecosystems. Whether a professional is tasked with analyzing massive spreadsheets, designing slide presentations, or drafting extensive corporate documents, the model is engineered to complete the work accurately. By deeply understanding the nuances of professional software, it delivers exactly what is requested with significantly less need for continuous user prompting.

A particularly groundbreaking feature introduced in the consumer facing ChatGPT is the innovative GPT-5.4 Thinking mode. This feature completely reimagines the dynamic between human and machine by providing an upfront, transparent plan of its intended thought process before generating the final output. Users have the unprecedented ability to monitor the planned trajectory of the artificial intelligence and seamlessly adjust the course mid response while the system is still actively working. This real time steering ensures that the final deliverable closely aligns with the exact needs of the user. Furthermore, this mode drastically enhances the capacity for deep web research, maintaining vital context over extended reasoning periods.

For software developers and enterprise users utilizing Codex and the API, the release introduces state of the art, native computer use capabilities. GPT-5.4 is the first general purpose model from OpenAI inherently designed to allow autonomous agents to operate desktop computers and execute complex workflows across entirely different applications. Supported by a massive one million token context window, these digital agents can meticulously plan, execute, and verify tasks over long horizons. This technological breakthrough means the artificial intelligence can visually navigate desktop interfaces, issue precise mouse commands, and manage administrative duties such as scheduling calendar events.

Despite these massive increases in capability, the new model is remarkably efficient. It stands as the most token efficient reasoning model OpenAI has ever produced, solving complex problems using significantly fewer tokens compared to its immediate predecessor, GPT-5.2. A major catalyst for this efficiency is the introduction of a new tool search feature within the API. Rather than loading thousands of tool definitions into the system memory for every single request, the model now accesses a lightweight directory. When a specific tool is required, it dynamically looks up the definition and incorporates it into the active conversation, cutting down token consumption for tool heavy operations.

Building upon its robust foundation, the model also shines brightly in software development and visual perception scenarios. A new fast mode in Codex significantly increases token velocity, allowing developers to maintain their creative flow while debugging or rapidly iterating on complex code structures. Visual perception has also seen massive upgrades, with the model now capable of processing incredibly high resolution images and parsing dense documents with minimal errors. An experimental feature named Playwright Interactive even allows the artificial intelligence to visually debug web applications or autonomously playtest interactive video games it is actively programming.

Finally, OpenAI has prioritized safety and factual reliability, making this release their most trustworthy model to date. Internal evaluations demonstrate a substantial decrease in hallucinations, with individual factual claims being significantly less likely to be false compared to previous generations. In rigorous professional benchmarks such as GDPval, which evaluates performance across numerous distinct occupations, the model successfully matched or exceeded the performance of human industry professionals in the vast majority of test cases. Industry leaders are already celebrating the ability of the model to deliver high quality, dependable results, cementing its role as a transformative asset for the modern workforce.