xAI releases Grok 4.5 as cheaper Opus-class alternative

At the International Conference on Machine Learning in Seoul, OpenAI senior researcher Noam Brown stated that GPT-5.6 would outperform human research interns for most tasks. This claim directly addresses CEO Sam Altman's October prediction that OpenAI would develop an AI-powered research intern by September 2026. The statement suggests the company is moving toward automating research roles, potentially reducing demand for human internships at the organization.

NVIDIA's Nemotron 3 Ultra model, tuned through LangChain's Deep Agents harness, achieved benchmark-leading performance on agentic AI tasks at one-tenth the inference cost of leading closed models. The optimization came through engineering the orchestration layer rather than retraining the model itself. Companies including Abridge, Amdocs, Box, and EY are already embedding specialized agents built on this stack into their platforms.

Chinese AI developer MiniMax is developing a 2.7-trillion parameter language model, which would be larger than any currently available Chinese AI model. The model could launch as early as Q3 2026, according to sources with knowledge of the plan. This represents a significant scaling effort by a Chinese AI company in a competitive market dominated by larger players.

Anthropic published research showing that Claude language models have spontaneously developed an internal structure called J-space that mirrors global workspace theory, a leading neuroscience model of human consciousness. Using a new mathematical technique called the Jacobian lens, researchers identified a privileged zone of internal activity where Claude holds concepts it can report on and reason with, surrounded by automatic processing it cannot access. The finding has already begun influencing how Anthropic monitors its AI systems for safety risks.