Particle.news

Download on the App Store

OpenAI Rolls Out GPT-5 With Key Enhancements While AI Agents Struggle Under Reliability Hurdles

GPT-5 cannot learn continuously despite its reduced hallucination rate

Image
The Amazon booth at an Agentic AI conference in Shanghai, July 2025 — when Amazon won't let AI Agents use its service.
Image

Overview

  • OpenAI began rolling out GPT-5 on August 7 to all ChatGPT users, making the new model available to free and subscription tiers.
  • It dynamically routes queries between a fast, nonreasoning version and a slower reasoning model, cutting hallucinations compared with GPT-4o.
  • The updated model expands its context window to 256,000 tokens, enabling sustained engagement with longer texts and complex prompts.
  • GPT-5 still cannot learn continuously after deployment, a capability OpenAI CEO Sam Altman highlighted as essential for true AGI.
  • Real-world trials of AI agents powered by similar models continue to see failure rates above 90%, prompting companies to impose strict operational guardrails and seek insurance against cascading errors.