Seven hours and fifty minutes. That's how long GPT-5.4 worked on a single task before delivering its answer, according to a user who shared their session log this week.
For context, when OpenAI first introduced deep research mode in early 2025, a "long" run meant maybe 5 to 10 minutes of the model browsing the web and synthesizing information. Extended thinking in reasoning models like o1 topped out at a few minutes of internal chain-of-thought. A nearly 8-hour continuous run is a different animal entirely - closer to handing a research assignment to a junior analyst than asking a chatbot a question.
The practical question is whether marathon sessions like this actually produce better results than breaking a problem into smaller pieces and prompting multiple times. In my experience with extended reasoning runs, there's a sweet spot. Longer isn't always better - models can go down rabbit holes, repeat themselves, or lose the thread of what they were originally asked. But for genuinely complex research tasks that require synthesizing dozens of sources or working through multi-step analysis, having a model that won't time out after a few minutes matters.
The compute cost of a single 8-hour run is also worth thinking about. OpenAI hasn't published per-minute pricing for extended reasoning sessions, but if you're on a Pro plan at $200/month, runs like this are where you start getting real value compared to the free or Plus tiers that cap reasoning time much more aggressively.
Whether 8-hour AI reasoning sessions become routine or remain edge cases will depend on how reliably they outperform shorter, iterative approaches. For now, the ceiling keeps rising.