OpenAI [Reuters=Yonhap News]

OpenAI, the developer of ChatGPT, has unveiled a new AI model that integrates knowledge work requiring reasoning and coding capabilities.

OpenAI introduced "GPT-5.4" and "GPT-5.4 Pro," improved versions of its top-tier AI models, on the 5th (local time).

A key feature of these models is their integration of previously separate models such as 'GPT-5.2 Thinking' and 'GPT-5.3-Codex' for specific domains.

Additionally, for the first time in its general-purpose models, it has incorporated a 'Computer Control' agent function that can directly operate a user's computer and perform complex tasks across multiple applications.

The processing capacity for a single input has also been increased to up to 1 million tokens (input/output units for AI models), enabling AI agents to formulate and execute long-term plans.

The GPT-5.4 Pro model achieved an 82% score on "GDPval," a performance benchmark evaluating knowledge work capabilities across 44 job categories, surpassing its predecessor's 70% level.

Furthermore, on the "BrowseComp" metric, which measures information retrieval capabilities, it scored 89.3%, higher than competitors Google Gemini 3.1 Pro (85.9%) and Anthropic Claude Opus 4.6 (84%).

In the "SWE-bench-plus-public" benchmark for coding abilities, GPT-5.4 received a score of 57.7%, which is higher than Gemini 3.1 Pro (54.2%).

However, OpenAI did not disclose the results for the "SWE-bench-verified" metric, which it has previously used as a standard for evaluating coding abilities. OpenAI explained that this metric's results were contaminated due to exposure in AI training data.

Claude Opus 4.6 had maintained the highest score on "SWE-bench-verified" until now.

The GPT-5.4 model, unveiled today, will be immediately available to ChatGPT premium subscribers. The previous model, 'GPT-5.2 Thinking,' will be maintained for three months and will be discontinued on June 5th.