The claude plugin eval command runs a plugin's eval suite against Claude Code, grades the output, then reruns each case without the plugin to show the difference, emitting scored JSON and HTML reports. MarkTechPost counts 6 grader types and a CI gate for skills; the release also adds /output-style and a Bash tool file-diff.
Anthropic is set to expand its revolving credit facility to $15 billion, up from a $2.5 billion five-year facility secured last year, per people familiar. Morgan Stanley is leading the process, clearing a hurdle before the AI firm's public IPO filing.
YuE2-3B, a 3B music generation model from m-a-p, takes lyrics plus a style prompt, writes a melody-and-chord plan, then renders it as a full song with vocals and accompaniment. The white-box workflow exposes the symbolic plan for editing; a ComfyUI node adds staged generation with an editable score/ABC workflow.
Ox Alpha appeared on OpenRouter August 20 as a free model with a 1M-token context window and text, image, and video input. A viral 10-task DeepSWE sample put it at 80% versus Claude Fable 5's 65% and GPT-5.6 Sol's 52%, but two full 113-task runs landed near 63%, roughly level with GPT-5.6 Sol.
Simon Willison had GPT-6 Astra build a Pluribus-themed Fabergé egg in Blender, a run that took 17m51s and produced several .blend files. Other users report Astra generating 3D racing environments from 8 office photos and rigging a quadruped from a reference video.
ARC Prize says ARC-AGI-4 will benchmark autonomous open-ended innovation, keeping the effort open-source as a shared research target. The announcement notes humans still significantly outperform AI at open-ended tasks.