Models & Releases 1
Max 5x gets $100 and Max 20x $200 monthly. Team pools $20 per Standard seat and $100 per Premium seat, capped at $500. Credits require a linked Console organization; interactive Claude Code and plan overage are excluded.
Tools & Demos 2
The SDK embeds policy-based containment into applications, with Windows process containers, Linux Bubblewrap and macOS Seatbelt plus additional backends. Filesystem, network and UI restrictions differ by backend; several VM options are experimental.
Thariq’s open-source extension reads browsing summaries and up to eight pages in the signed-in browser, then sends cleaned content to Anthropic to create a fresh homepage. Private page content can be included; each build consumes the user’s API balance.
AI Agents 2
An 8B agent links image/video regions, entities, retrieved sources and answer steps through a visually grounded evidence graph. Authors report gains of 20.2 and 17.6 points over tool-enabled Qwen3-VL-8B on two newly introduced benchmarks.
An evidence-grounded behavior graph groups source-linked actions to help monitors find consequential decisions and check requirement alignment. Experiments across eight models improve decision identification and evidence localization in most settings without additional training.
Lab Updates 2
Brice Ménard combined UV surveys and used other wavelengths to predict missing regions, about one third of the sky. The published map distinguishes measurements from predictions and includes uncertainty layers; it is not a new direct observation of every region.
The OpenAI customer story describes matching model capability to task complexity, restricting Fast mode and assigning budgets by business stage. Estimated daily cost fell about 65% versus its earlier setup; feature-level customer-value measurement is still being built.
Podcasts 1
In Latent Space, Liam Fedus and Ekin Doğuş Çubuk explain why automated synthesis, measurement and learning from failed experiments must accompany capable models. A complete publisher transcript was collected; proposed discoveries and business plans remain the guests’ claims.
TLDR: Scientific progress needs experiments that can falsify predictions. Failed and noisy measurements are valuable training data. Evaluate physical AI by reproducible experimental outcomes, not fluent proposals. Chapters: physical-world RL 02:49; discovery loop 09:17; experimental ground truth 30:22; negative results 53:16; deployment 01:10:45. Liam Fedus: “you can’t just think your way to a solution” [00:00:29]. Ekin Doğuş Çubuk: “What’s really hard is to synthesize them” [01:21:23].
Industry 2
The November 12 update clarifies deceptive campaigns, surveillance, weapons software, high-risk decisions and autonomous hardware. An extreme-abuse clause excludes ordinary frustration, disagreement, dark fiction and research; ending the interaction remains its primary enforcement mechanism.
The consortium combines funding, data, compute and measurement infrastructure for open predictive-biology resources. The total includes more than $500M of prior NIH-supported investment, not just new cash; DeepMind, Isomorphic Labs and Meta contribute $300M collectively.
HF Trending Papers 2
Request-level routing logic feeds asynchronous per-model subservers with delayed batching. Authors report 2.01–64.15× higher decoding throughput against tested systems across different routing workloads; this range is not a speedup guarantee for every ordinary LLM request.
The world model predicts videos conditioned on robot actions, calibrates trajectories and learns from modified action scenarios. Authors report human-rated interaction defects falling from 48.12% to 6.25% on AgiBot; this evaluates generated predictions, not deployed robot safety.