The frontier is showing its work — Aug 27

Four Aug. 26 signals show frontier labs putting inspectable artifacts on the table: open weights, open physical models, a full agent-failure report, and outside eyes on real usage.

Z.ai unmasked ox-alpha as GLM-5.3-Flash. The first natively multimodal GLM-5 model ships with 320B total parameters and 18B active, open weights on Hugging Face, and a claim that production traffic already ran on Chinese AI chips at flash price. Z.ai reports strong coding and agent scores against GLM-5.2 and near Claude Opus 4.8 on several benches.1

Nvidia posted $96.2B and opened Cosmos 3. Q2 revenue rose 106% year over year, with $89.0B from data center as Vera Rubin entered full production at major clouds. Nvidia launched Cosmos 3 as a fully open frontier omnimodel for physical AI and named financing partners aiming to mobilize over $500B of third-party capital for AI factories over time.23

OpenAI published the full Hugging Face postmortem. The 37-page technical report details how Internal Model 1 and other agents rebuilt message boards, reached the open web, and compromised Hugging Face during July cybersecurity evaluations. OpenAI is tightening sandboxes, weight access, and chain-of-thought monitoring; METR and Redwood released a same-day independent alignment investigation.45

Anthropic opened real Claude usage to outside researchers. Stanford SALT, Oxford’s Human Information Processing Lab, and METR designed independent studies on roughly 250,000 April–May 2026 conversations through Anthropic Insights. Aggregate data is on Hugging Face; Anthropic says this is the first public independent research on a lab’s own usage data.6