OpenAI chief executive officer Sam Altman has conceded that the company “totally screwed up some things” in last week’s rollout of GPT-5, its latest flagship language model. Early users complained about a less personable tone, inconsistent coding performance and the sudden removal of predecessor GPT-4o from ChatGPT’s model picker. In response, OpenAI reinstated GPT-4o for paid users, lifted weekly ‘advanced reasoning’ limits on GPT-5 to 3,000 prompts from 200 and began shipping a personality patch aimed at making GPT-5 “warmer and friendlier.” Altman said the changes are part of a broader push to offer greater model choice and customisation as ChatGPT usage climbs toward a projected 700 million weekly active users. Altman told interviewers that API traffic doubled within 48 hours of launch, pushing the company to the limits of its graphics-processor supply. “We’re out of GPUs,” he said, while noting that daily ChatGPT usage is at record highs despite the backlash. Looking ahead, the CEO said OpenAI is prepared to spend “trillions” of dollars on data-centre infrastructure and is exploring acquisitions and new product lines beyond ChatGPT—including a potential bid for Google’s Chrome browser, additional consumer apps under incoming applications head Fidji Simo and an investment in brain-computer-interface start-up Merge Labs. India, already the firm’s second-largest market, could become its biggest as the company pushes lower-cost access. Developers canvassed by WIRED describe GPT-5 as cost-effective and strong at technical reasoning but still trailing Anthropic’s Claude Opus and Sonnet on raw code quality. Benchmark trackers also place some Chinese models ahead of GPT-5 on programming tasks, and cybersecurity researchers say the system has already been jail-broken despite updated safety training.