Aug/2025 - OpenAI GPT-5 - LifeArchitect.ai LIVESTREAM

Quick Overview

GPT-5 represents an iterative step for OpenAI, primarily focused on cost reduction and serving a larger user base, rather than a groundbreaking leap in capability compared to GPT-4.1 or GPT-4 Turbo. While it aims to replace GPT-4o, its performance on benchmarks like the "OpenAI proof question and answer" is notably low, scoring only 2%. The model's architecture appears to be a 'broken router' pinging between six different models, and its effectiveness is questioned, with some users noting it seems 'way dumber' when the 'order switcher' is out of commission. Despite initial concerns about its capabilities and lack of significant advancement over previous models, OpenAI is reportedly considering bringing back GPT-4o for plus users.

Key Points: GPT-5 is positioned as an iterative update focused on cost reduction and serving OpenAI's massive user base, rather than a significant leap in capability over models like GPT-4.1 or GPT-4o. The model's architecture is described as a 'broken router' that pings between six different models, with its overall logic for this routing being unknown. Performance on specialized benchmarks is underwhelming, with GPT-5 thinking scoring only 2% on the "OpenAI proof question and answer" benchmark, a task that requires significant time for a team at OpenAI to solve. Users have reported that GPT-5 can seem 'way dumber' when certain internal systems, like the 'order switcher', are not functioning correctly. Despite the perceived lack of significant improvement over GPT-4o or GPT-4.1, OpenAI is considering reintroducing GPT-4o for 'plus' users due to user feedback. The transcript highlights concerns about hallucination rates, noting that a benchmark previously used to track this (person QA) was removed from the GPT-5 paper, with estimates suggesting GPT-5's hallucination rate might have doubled compared to previous bests. The data used for GPT-5 training is estimated to be over 100 trillion tokens, with potentially half being synthetically generated data, and the cutoff date remains September/October 2024, similar to GPT-3.5 Turbo.