Cur8 — Monday, September 21, 2026
Fable 5.2 challenges GPT-6, while Kev introduces compact decision models based on Qwen3.5. OpenAI advances AI agents with institutional memory, and Google unveils AX, an agentic orchestrator. Meanwhile, YouTube highlights legal and agentic AI trends, including Meta's Muse and persistent automation.
New model releases
Anthropic's Fable 5.2 builds a Brawl Stars clone in 1 hour. Anthropic rerouted users to 5.2 without notice.
Kev-4B and Kev-9B models (based on Qwen3.5) achieve 0.837–0.852 accuracy on new sources, with Kev-9B showing 0.852 test accuracy, outperforming previous Kev versions. They support multi-question inference with shared input.
OpenAI news
V7 uses GPT-5.6 to create institutional memory for AI agents, enabling them to access and use company files for complex, source-linked tasks.
Claude Code dominates a major company's workflow, generating specs, code, and documentation. Engineers work 12–13 hours daily, relying on Claude to "press enter."
Big cloud & vendor AI news
Google's AX enables billions of isolated agentic tasks per cluster with declarative control, sandboxing, workspace setup, and model integration. It scales stateful actors with sub-second suspend/resume, reducing idle costs and supporting generative AI workflows.
YouTube trending AI content
US judge declines to break up Google but mandates open access to its ad exchange, citing enforcement challenges and rapid tech evolution. Judge Leone Brinkma emphasizes lack of trust in Google's compliance.
U.S.-Greenland security deal announced; unclear if it changes current arrangements.