General · 6 March 2026
GPT-5.4 scores 75% on computer benchmark
OpenAI released its new GPT-5.4 model for ChatGPT, Codex, and developer API users. It starts rolling out today for Plus, Team, and Pro customers. The model natively controls computers, making it the first general-purpose OpenAI tool with this ability. It achieved a 75.0% score on the OSWorld-Verified benchmark, exceeding the 72.4% human baseline. It is 33% less likely to produce false factual claims than GPT-5.2, and financial spreadsheet modelling scores improved to 87.5% from 68.4%.
Reported by Tom's Guide · How we write briefs