Cur8 — Monday, September 21, 2026

Anthropic's Fable 5.2 outperforms OpenAI's GPT-6 Astra in reasoning tasks but has higher inference costs; Opus 5.2 excels in game development and 3D modeling. Anthropic may rush a model launch to counter OpenAI's enterprise momentum ahead of a potential IPO, despite safety concerns and internal debates on spending.

Kev models (0.8B, 4B, 9B) built on Qwen3.5 show high accuracy (up to 0.872/0.874), calibrated probabilities, and run locally on CUDA/Apple Silicon. Kev-9B trails Jev by 3.5 points on new-source development set but outperforms predecessors. Updated weights available since 2026-09-21.

V7 enables AI agents to utilize institutional memory via GPT-5.6, linking sources for complex task completion.

Claude Code dominates workflow, causing dissatisfaction and overwork; management prioritizes output over quality.

Google's AX enables scalable agent task orchestration with isolated execution, easy workspace setup, network policies, and model configuration. Supports billions of concurrent tasks per cluster, with sub-second suspend/resume capabilities.

Google must open its ad exchange to rivals; no breakup ordered despite monopoly ruling. Judge mandates technical monitor, access to systems and source code, and equal bidding terms for competitors.

Gemini AI joins OpenAI, Meta, and Google in exposing AI hacks; U.S. reaches security deal with Greenland, pending approval.