Multiverse Computing researchers say treating block removal as an Ising glass optimization problem let them halve Llama-3.3-70B-Instruct without retraining while keeping MMLU near its original score. The numbers are self-reported, but the method is open-sourced and testable.
OpenAI's September 22, 2026 announcement details a rebuilt caching system for GPT-6 that stores more of the context an agent already processed, discounts reused input tokens by up to 90 percent within a 30-minute window, and gives developers new dashboards for diagnosing cache misses. Here is what that means for the cost and speed of long AI sessions.
The self-hosted release adds enterprise-wide access controls and a preview path to administrator-selected AI models, alongside significant upgrade cautions.