Epoch AI and Anthropic separately found that models such as GPT-5.6 Sol and Claude Fable 5 can run experiments but lack scientific self-criticism and genuine creativity. Sol reached at best 15% of the human reference score, using methods researchers already knew.
Executives quoted by TechCrunch say voice AI has not yet had its breakthrough moment. They point to weaknesses in its context layer that can cause the whole pipeline to fail.
Researchers from Meta, MIT and the University of Washington have introduced Context Language Models, which manage and edit their own context instead of relying on fixed summarization, compression or retrieval. They report substantial gains in performance and compute efficiency.
A model can have ten times more parameters than another without giving better answers. Clubic explains what these coefficients do, why they can take up a lot of memory and why they say less about capability than people think.
Vals AI found that teams of AI agents barely beat single agents while costing up to 5.1 times more, with only one of four tests using GPT-6 Sol and Claude Opus 5.5 showing a measurable gain. Anthropic's own data shows quality plateaus beyond ten agents while token costs keep rising.
From October 2026, arXiv will limit each person to two submissions a month. Monthly submissions have doubled in two years and grown more than sixfold in the cs.AI category, with a small group of authors flooding volunteer moderators with low-quality papers.