codex is now built for long tasks. but not for users yet?
we called it yesterday: openai is optimizing agents for long-horizon thinking. now thibault sottiaux, codex team lead at openai, confirms it – codex can run for a really long time
reminder: sam altman said chatgpt 5.5 will consume fewer tokens per answer or action. so the only playbook left to grow revenue is:
1. teach ai to handle the hardest tasks
2. ship it to users
3. tell them ai can now do the work they hate doing
thehype tested new codex with chatgpt 5.5 medium – asked it to generate an infinite day-to-night landscape animation
after 2 mins 15 secs of "thinking," it produced 490 lines of code. the trees were flying over the river...
do we really need longer tasks when the short ones still don't work?
With some small tweaks, Codex can work for days on hard tasks. We will release some changes to make this easier to use for everyone.
— Tibo (@thsottiaux) April 29, 2026
What’s the hardest task you’ve seen GPT-5.5 succeed at?
Nick Trenkler