theinformation.com
|
ksl
|
|
OpenAI is preparing GPT-5.4 with a one-million-token context window – more than doubling GPT-5.2’s 400K limit – and a new “extreme” thinking mode that throws significantly more compute at hard problems. The extreme mode is aimed squarely at researchers, not general users, which makes it functionally a premium reasoning tier. OpenAI also says the model will be more reliable on multi-hour tasks, a direct nod to its Codex agent where long-running sessions tend to accumulate errors. The million-token context catches OpenAI up to where Google and Anthropic already sit. No release date yet, though OpenAI teased it would arrive soon. The competitive pressure to ship reasoning upgrades has turned model releases into near-monthly events across all three frontrunner labs.
