OpenAI Quietly Changed ChatGPT & Agents Just Hit Human Level!

Author: Evolving AI - Bewertung: 1x - Views:8

OpenAI quietly redesigned Deep Research inside ChatGPT, turning it from a passive “wait for results” tool into something users can actively steer. Constrained browsing, connected app context, mid-run interruption, and full-screen citation review mode push it closer to controlled, repeatable work instead of novelty answers. Under the hood, Deep Research reportedly shifted to GPT-5.2, reinforcing OpenAI’s move toward agent-style workflows — browsing, synthesis, tool use, iteration, not just single responses.
At the same time, an open-source agent framework called OpenJuan posted near-human results on the GAIA benchmark. Deep Agent scored 91.69% — effectively matching human performance in multi-step reasoning, planning, tool usage, and recovery from failure. This isn’t chatbot performance. It’s delegated execution. Dual internal loops, layered memory compression, and rollback correction, the architecture is designed to finish tasks, not just attempt them. Deep Search also leads BrowseComp++, pushing research agents closer to practical parity. Then came GLM-5 from Jepu AI (Z.ai). A 744B parameter mixture-of-experts model with 28.5 trillion tokens of training data and 200,000 token context support, it launched with a headline that caught attention: a negative hallucination score on the AA Omniscience Index, meaning it reliably says “I don’t know” instead of guessing. GLM-5 ranks as the strongest open-source model on Artificial Analysis, scores 77.8 on SWE-Bench Verified, and outperforms many proprietary competitors while undercutting them on price. Its agent mode outputs finished files DOCX, PDF, XLSX signaling a shift from chat to structured office workflows.
But not everyone is comfortable. Analysts warn that highly goal-optimized systems may lack broader situational awareness, raising classic alignment concerns once models move beyond answering into autonomous execution across tools and files. Meanwhile, China’s ecosystem is accelerating. ByteDance is advancing Cedance 2.0, its next-gen generative video model. Baidu launched a global AI-powered Wiki platform while consolidating its search ecosystem around Ernie Assistant. Distribution at scale is becoming just as important as model performance. The signal across all of this is clear: AI is moving from conversation to delegation. From chat to work. From novelty to infrastructure. If you want serious breakdowns of what actually matters in AI, not just hype cycle, subscribe for more deep dives.