New GPT-5.4 clobbers humans on pro-level work in OpenAI's tests - by 83%

GPT-5.4 is also more reliable, producing 18% fewer errors and 33% fewer false claims than GPT-5.2, according to OpenAI.