I’ve spent about five hours with GPT-5. Not the press release. Not the benchmark charts. The actual model. Long enough to move past the “wow” phase and into the “what’s really new here?” phase. Here’s what I noticed.
The company says GPT-5 can hold much longer context. I started by asking it to draft a workshop agenda for a client. After hours of unrelated work, I told it to “make it a half-day.” It remembered exactly what we had been working on and produced a new version without asking for the original; previous models usually needed me to paste the whole thing back in. OpenAI credits this to an expanded context window exceeding 200k tokens with improved attention sparsity, enabling more efficient retrieval of relevant prior turns without full recomputation.
OpenAI also claims improved reasoning. I asked for a three-month marketing AI pilot plan. It delivered one, then when I cut the budget, it reworked the entire thing without starting over. Earlier models tended to lose the thread and rebuild from scratch. OpenAI says this comes from multi-step reasoning modules with persistent latent state tracking, which allow the model to iteratively refine outputs without regenerating the full reasoning chain.
Speed has improved, too, without losing detail. I requested a side-by-side comparison of three AI image-generation platforms, including pricing and standout features. The results came back in seconds, with the same thoroughness I’d expect from a slower GPT-4o run. Older models could be quick, but usually only if you asked for less depth. OpenAI attributes the gain to a hybrid mixture-of-experts (MoE) architecture with dynamic expert routing, plus speculative decoding that pre-computes high-confidence token sequences before final confirmation.
It also takes more initiative. For a panel on AI governance, I asked for talking points. It built the list, then suggested a slide order and even wrote an opening joke that was pretty good. Earlier models might have asked if I wanted slides or a script before doing anything extra, and the jokes would have been terrible. OpenAI says this comes from integrating agentic planning layers that break down high-level goals into sub-tasks, with a policy network selecting the most probable next actions without explicit prompts.
Finally, GPT-5 communicates more naturally. When I asked about an obscure regulation, it simply said, “I don’t have reliable information on that,” and moved on. No filler, no disclaimers. OpenAI says this comes from fine-tuning with human preference alignment data and adversarial prompt filtering to cut low-value meta-commentary while preserving directness and factual calibration.
I do not want to anthropomorphize an AI model, but five hours with GPT-5 felt more like working with someone than working with something. It kept track of the work, adjusted to changes without a fuss, worked quickly, and offered contextually useful next steps.
My unscientific first impression is that this model can keep up with me. I’ll use it exclusively today and see if I feel the same way tomorrow.
As always your thoughts and comments are both welcome and encouraged. -s
Shelly Palmer is the Professor of Advanced Media in Residence at Syracuse University’s S.I. Newhouse School of Public Communications and CEO of The Palmer Group, a consulting practice that helps Fortune 500 companies with technology, media and marketing. Named LinkedIn’s “Top Voice in Technology,” he covers tech and business for Good Day New York, is a regular commentator on CNN and writes a popular daily business blog. He's a bestselling author, and the creator of the popular, free online course, Generative AI for Execs. Follow @shellypalmer or visit shellypalmer.com.
