OpenAI’s next flagship model, GPT-5, is nearing release but is not expected to deliver the dramatic performance jump seen between GPT-3 in 2020 and GPT-4 in 2023, according to reporting by The Information. People briefed on the project said the company’s earlier attempt, an iteration internally codenamed “Orion” and once slated to be marketed as GPT-5, fell short of targets and was re-labelled GPT-4.5 before being shelved. The outlet says engineers struggled with a dwindling supply of high-quality web data for pre-training and found that reinforcement-learning techniques that succeeded on smaller systems did not scale reliably. Those issues, combined with rising computational costs, have slowed overall progress and forced the team to look for alternate approaches such as “universal verifiers” to boost accuracy. Despite the setbacks, GPT-5 is still expected to show meaningful, if incremental, gains in practical tasks such as computer programming, mathematics and autonomous agent control. Speaking on a podcast recorded 30 July, OpenAI Chief Executive Officer Sam Altman acknowledged the system’s power and said aspects of testing the model left him “nervous,” underscoring broader industry concerns about increasingly capable AI.