Native Computer Control and Multimodal Perception
GPT-5.4 is OpenAI’s first general-purpose model with native computer control capabilities, enabling it to perform tasks across different software and web pages by recognizing graphical user interfaces (UIs) and directly manipulating the mouse and keyboard. On the OSWorld-Verified benchmark, which measures an AI agent’s ability to operate an OS, GPT-5.4 achieved a success rate of 75.0%, surpassing the previous record of 72.7%. This capability is powered by its enhanced visual perception, with the model scoring 81.2% on the MMMU-Pro test. It also introduces “original” and “high” image input modes that support higher resolutions, improving its analytical accuracy for complex documents and interfaces.
Deep Fusion of Reasoning and Programming
The model achieves a deep integration of reasoning and programming abilities, fully inheriting the code generation DNA of GPT-5.3-Codex. On the SWE-Bench Pro test, which focuses on real-world software engineering tasks, GPT-5.4 achieved an accuracy of 57.7%. Simultaneously, it ranked first on benchmarks like FrontierMath for mathematical reasoning and ARC-AGI-2 for abstract reasoning, scoring 83.3% on the latter. This integration aims to provide an all-in-one development and analysis tool that eliminates the need to switch between different models, while also improving token efficiency and processing speed.
Improved Practicality and Factual Accuracy
To reduce AI hallucinations, GPT-5.4 has been optimized for factuality. According to internal testing, its probability of making erroneous claims has been reduced by 33% compared to GPT-5.2. On the GDPval benchmark, which simulates professional knowledge work, the model achieved an 83.0% win rate, demonstrating its professional-grade capabilities in handling office tasks like spreadsheets and presentations. Furthermore, the newly introduced “Tool Search” feature can reduce total token consumption in complex tool-use scenarios by 47% while maintaining accuracy, significantly boosting the operational efficiency of AI agents.
Model Versions and Pricing Strategy
OpenAI has released several versions targeting different users. On the ChatGPT client, GPT-5.4 Thinking and GPT-5.4 Pro are available. Developers can access GPT-5.4 and its high-speed version, GPT-5.4 fast, via the API. In terms of pricing, the standard GPT-5.4 API is priced at $2.50 per million input tokens and $15 per million output tokens. The more powerful Pro version is priced at $30 per million input tokens and $180 per million output tokens.
