OpenAI has introduced GPT-5.4, a new AI model that combines recent improvements in reasoning, coding and task automation into a single system. The model builds on the capabilities of earlier versions and it also improves on the coding strengths seen in GPT-5.3-Codex. Essentially, the aim is to make the model better at handling practical work, like when writing code, analysing spreadsheets, creating presentations and even when working across different software tools.

In ChatGPT, GPT-5.4 Thinking can now outline a plan before generating a full response. This allows users to review the approach and then adjust instructions while the response is still being prepared. The update further improves how the model handles web research, especially for complex or highly specific queries. Furthermore, it can maintain better context during longer conversations.

For developers, GPT-5.4 introduces built-in computer-use capabilities. This means AI agents can interact with software and websites to complete multi-step tasks across different applications. The model supports up to one million tokens of context, allowing it to handle longer workflows and larger datasets. It also introduces a tool search feature that helps the model identify and use the right tools more efficiently without loading all tool definitions into every request, which can reduce costs and speed up responses.

The model is also designed to use fewer tokens when reasoning through problems compared to GPT-5.2, which helps in improving efficiency as well as response speed. In internal evaluations, GPT-5.4 showed improvements in professional knowledge work tasks as per OpenAI. On the GDPval benchmark, which measures performance across 44 occupations, the model matched or outperformed human professionals in 83 percent of comparisons, compared with around 71 percent for GPT-5.2.

Also Read: ChatGPT vs Gemini vs Grok Free Plans: Comparison and Which One Would Suit You

OpenAI also focused on improving tasks commonly used in office environments. On an internal spreadsheet modelling benchmark similar to work done by junior investment banking analysts, GPT-5.4 scored 87.3 percent compared with 68.4 percent for GPT-5.2. In presentation-related tasks, human evaluators preferred presentations generated by GPT-5.4 in 68 percent of cases due to improved visual layout and design.

Accuracy has also been improved. In tests based on prompts where users previously flagged factual mistakes, GPT-5.4 produced responses with 33 percent fewer incorrect claims and 18 percent fewer overall factual errors compared with GPT-5.2.

Next, the model also improves visual understanding, particularly when analysing high-resolution images. OpenAI says GPT-5.4 can process images with up to 10.24 million pixels when using its highest detail setting, which helps with tasks such as image analysis and interface interactions.

For developers building AI agents, GPT-5.4 is designed to work well with tools and external systems. It can further write code to control computers through automation libraries or even interact with software through keyboard and mouse commands. Developers can further also adjust how the model behaves through configuration settings and confirmation policies.

The model has also improved at web browsing tasks. On the BrowseComp benchmark, which measures how well AI systems can search the web for difficult-to-find information, GPT-5.4 improved performance by 17 percentage points compared with GPT-5.2.

GPT-5.4 is rolling out starting today across ChatGPT, Codex and the API. In ChatGPT, GPT-5.4 Thinking is now available to Plus, Team and Pro users, replacing GPT-5.2 Thinking. The older model will remain available in the Legacy Models section for three months before being retired on June 5, 2026. Developers can access the model through the API as "gpt-5.4", with a higher-performance version called "gpt-5.4-pro" also available.