GPT-5 Arrives (to a lukewarm reception)
After much anticipation and fanfare, OpenAI has unveiled GPT‑5, an AI system that (according to OpenAI) performs at a “PhD” level across coding, writing, math, and other tasks
GPT-5 is a “system,” more than an LLM, because GPT-5 switches between multiple GPT-5 models. GPT-5 automatically switches between fast and deep reasoning modes, enabling more “thinking” in reasoning mode when necessary.
Premium subscribers can directly select the “GPT-5 Thinking” reasoning model:
OpenAI claims that GPT-5 generates fewer hallucinations, and is more honest, safer, and stronger than previous models.
Free and paid ChatGPT users have immediate access, with paid tiers enjoying extended usage and capabilities. OpenAI also offers “GPT‑5 pro” for ultra‑complex queries.
Tom’s Take: Noted AI researcher and Wharton professor Ethan Mollick had early access to GPT-5 and calls it “a big deal.” Mollick likes how GPT-5 decides how much reasoning “effort” to apply. Few others share his enthusiasm. AI analyst and researcher Michael Spencer calls GPT-5’s launch a “failure” and scientist and author Gary Marcus calls GPT-5 “overdue, overhyped and underwhelming.”
Put me in the “underwhelmed” category. After testing GPT-5, GTP-5 Thinking and Deep Research for a few hours, I noticed little improvement in the quality of responses. I also saw little evidence of more reasoning at work than in previous models. Responses were indeed fast, but GTP-5 didn’t follow directions several times.
Yet, even if GPT-5 only represents incremental progress in performance, I'll be satisfied if it can significantly reduce hallucinations. OpenAI claims GPT-5 commits 26% fewer hallucinations than GPT-4o, especially in the reasoning (“thinking”) variant. If that’s true, that’s a considerable improvement and would save me many hours checking its responses. Fingers crossed.
See related posts from Tom Daccord’s free newsletter:
