Headlines

1. Anthropic Launches Claude Opus 5, Opening a New Era of Verifiable Agentic Reasoning
Claude Opus 5 introduces a “director-of-search” deliberation routing system that dynamically switches between two-second rapid responses and four-minute deep reasoning tasks, setting new records of 67.2% on GAIA-2 and 48.9% on ARC-AGI-2. The model emits structured evidence logs for every claim, making it the first frontier LLM designed for cryptographic auditability in enterprise agent workflows.
Source: Anthropic press release; The Information

2. Nvidia Begins Volume Shipments of Rubin Ultra GPUs to Hyperscalers
Nvidia’s Rubin Ultra R100 parts started shipping to Microsoft Azure and Oracle Cloud Infrastructure this week, delivering 2.2 exaflops per rack in FP8 sparse configurations—an 11.2x performance-per-rack gain over the Hopper generation. Commercial services built on the new part are expected to reach general availability in November, and Nvidia says over 2 million units are now committed to data center builders.
Source: Reuters; Nvidia Newsroom

Model Releases

3. Google DeepMind Debuts Gemini 3 Pro with 2-Million-Token Hierarchical Mixture-of-Agents
Gemini 3 Pro replaces the unified decoder with a “senior/junior” mixture-of-agents architecture that partitions reasoning across 2-million-token contexts without losing global coherence, achieving a 1724 average Elo on the LMArena leaderboard with a 1.5-second time-to-first-token in streaming mode. A companion Agentic API allows developers to call research, code, and web subtasks as modular functions with built-in rollback.
Source: Google DeepMind blog; The Verge

4. Meta Releases Llama 5 175B as Open Weights with Verifiable Tool-Calling Contracts
The latest Llama 5 checkpoint adds a native “verify-at-inference” loop that executes tool calls, monitors their effects, and rolls back erroneous actions, scoring 78.1% on SWE-bench Verified—within 1.3 points of GPT-5.5-mini’s 79.4%. Apache 2.0 weights are now live on Hugging Face, with 27B and 70B vision-language variants scheduled to follow next month.
Source: Meta AI; Hugging Face

Industry

5. Microsoft and Constellation Break Ground on 3-GW Nuclear-Powered AI Campus in Pennsylvania
Project “Crimson Ridge” will combine the existing Susquehanna nuclear facility expansion zone with a 2.1-GW advanced pressurized water reactor and on-site small modular reactors, representing an $18.7 billion investment to power Microsoft’s Eastern US AI training workloads. The first phase is slated to come online in 2029, and Microsoft says the campus will be its largest single-tenant compute facility supporting OpenAI-model training.
Source: Bloomberg; The Wall Street Journal

6. OpenAI Signs 1-GW Geothermal Deal with Fervo Energy for Texas “Altair” Training Cluster
OpenAI executed a 20-year power purchase agreement with Fervo Energy to supply the new “Altair” training megacluster in the Permian Basin, where Fervo will deploy 400 geothermal wells that can ramp output in 30 seconds to stabilize intermittent wind and solar. The deal gives OpenAI 1 GW of firm, carbon-free power at a projected total cost 14% lower than comparable gas-peaker arrangements.
Source: Reuters; Fervo Energy

Research Highlights

7. DeepMind’s Chronos-X World Model Learns Causal Video Primitives from 1.5B Hours of Sensor Data
The 400B-parameter Chronos-X model predicts 10-second forward video rollouts with 79% top-1 scene-change accuracy and demonstrates zero-shot sim-to-real transfer on the ALOHA robotics stack for long-horizon manipulation tasks. DeepMind researchers say the model will form the core of its upcoming “Lattice” robotics foundation system.
Source: arXiv; Nature Machine Intelligence

8. Stanford and OpenAI Introduce Landmark Caching to Cut Agent RL Compute by 23x
A new paper presents Temporal Landmark Cache (TLC), which reuses successful subgoal trajectories across 15 million agent rollout episodes, reducing repeated environment steps by 91% with no measurable loss in HELM-agent composite scores. The technique is already integrated into the open-source gym-3 agent rollout library and could materially lower the energy cost of large-scale agent reinforcement learning.
Source: arXiv; Stanford HAI

Compiled from multiple media reports.