OpenAI launches GPT-5.4 with computer vision, tool use enhancements – News

OpenAI Group PBC today launched a new large language model that it says is more adept at automating work tasks than its earlier algorithms.

GPT-5.4 is available in ChatGPT, the Codex programming tool and OpenAI’s application programming interface.

The company bills API users based on the number of tokens that its LLMs process while generating a prompt response. A token is a unit of data that comprises a few letters or characters. OpenAI says that GPT-5.4 uses “significantly” fewer tokens than GPT-5.2, which debuted in December. Reducing a model’s token use drives down inference computing costs. 

OpenAI says that its new model can also reduce customers’ inference bills in other ways. 

Applications built on OpenAI’s API often rely on external programs, or tools, to complete tasks. Until now, developers had to prepare a detailed list of the tools that their applications use and include it in their API requests. A tool list can increase the size of API requests by thousands of tokens, which drives up inference costs. 

GPT-5.4 makes the workflow more efficient. According to OpenAI, a new search engine enables the model to automatically find the tools that an application requires to perform a given task. That avoids the need to upload detailed tool lists, which reduces prompt sizes and inference costs.

The new model can ingest requests with up to 1 million tokens. Compared with its predecessor, the model is significantly better at processing prompts that contain images. Developers can upload images that contain more than 10 million pixels without having to compress them, which prevents the loss of potentially important details.

The upgraded vision capabilities make GPT-5.4 more adept at computer use, or the task of interacting with applications via their user interfaces. OpenAI evaluated the model using a popular computer use benchmark called OSWorld-Verified. It set an industry record with a score of 75%, which is higher than both GP-5.2’s result and the 72.4% typically achieved by human testers.

The model also outperforms its predecessor in other areas. GPT-5.4 achieved a mean score of 87.3% on a spreadsheet analysis benchmark created by OpenAI, a more than 8% improvement over GPT-5.2. The former LLM is also better at preparing presentations, using browsers to perform online research and answering science questions.

GPT-5.4 is available through OpenAI’s API for $2.5 per million input tokens and $12 per million output tokens. Users with advanced requirements can access an enhanced edition of the model, GPT-5.4 Pro, that OpenAI says is designed to provide “maximum performance on complex tasks.” The enhanced edition is also available in ChatGPT alongside the standard version of the model.

Image: OpenAI