GPT-5: The Case of the Missing Agent (secondthoughts.ai)

OpenAI’s release of GPT-5 marks another leap in language model capabilities, boasting significant improvements over GPT-4 in speed, cost-efficiency, context window size (up to 400,000 tokens), and complex reasoning skills. GPT-5 shines in specialized benchmarks, such as solving 65% of problems on the challenging SWE-Bench Verified coding test, a massive jump from GPT-4 Turbo’s 2.8%. These advancements demonstrate improvements in multi-step reasoning, tool use, and handling extensive contextual information, positioning GPT-5 as a powerful foundation model for AI applications. However, despite these impressive internal gains, GPT-5—and other contemporary models like Anthropic’s Claude 4.1 and Google’s Gemini 2.5—still fall short as truly agentic AI capable of robust, autonomous operation in the messy real world. Experiments with AI agents managing tasks such as running a mini-store or operating in continuous goal-driven environments reveal profound limitations: hallucinated facts, failure to learn from mistakes, poor long-term planning, and losing track of their own identity as software entities rather than physical agents. GPT-5’s attempt to play Minesweeper highlights its inability to accurately perceive visual information, leading to repetitive and ultimately fruitless behavior. This gap underscores a critical distinction in AI progress: while models like GPT-5 excel at expanded reasoning and knowledge-based tasks, creating autonomous agents that can flexibly and reliably navigate dynamic real-world environments remains an elusive challenge. The hype around “agentic AI” often overlooks these fundamental limitations, suggesting that the journey toward genuine autonomous AI assistants with long-term goal management is still in its early stages.