It is reported that Anthropic has signed a $10 billion, six-year cloud capacity procurement agreement with infrastructure startup Volta. Data centers will be built in Norway using NVIDIA Vera Rubin systems, reserving long-term compute capacity for Anthropic's training and inference workloads. This arrangement binds multi-year capacity, dedicated campuses, and next-generation chips, indicating that frontier model companies are securing supply through long-term contracts rather than renting GPUs on-demand. The materials do not disclose the activation timeline, total capacity, procurement schedule, or payment structure; execution details remain subject to official announcements from both parties.

Google announced a restructuring of its AI leadership: Demis Hassabis will transition to Chairman of DeepMind and Chief Scientist at Alphabet, while Koray Kavukcuoglu is promoted to Senior Vice President at DeepMind, overseeing Gemini models, frontier research, and applications, reporting directly to Sundar Pichai. Jeff Dean and Sanjay Ghemawat will depart to launch an AI research nonprofit, with Google serving as founding investor and cloud partner.

Meta AI has released the beta version of Muse Code, a terminal-based coding agent powered by Muse Spark 1.2. The asynchronous background agent persists throughout sessions, with local append-only event logs recording model calls, tool executions, approvals, and code changes—enabling precise replay and crash recovery. The model and agent framework are jointly trained for long-horizon, repository-level tasks, evaluated across Terminal-Bench 2.1, DeepSWE v1.1, and 440 internal tasks. In one case, the agent ran continuously for 24 hours on an NVIDIA Hopper GPU, making over 1,000 tool calls while optimizing Triton kernels.

Tsinghua AIR and Berkeley BAIR have introduced ODEWorld, a continuous-time embodied world model that learns latent velocity fields via PT-Flow, moving beyond fixed-frame mappings to generate frames at arbitrary rates, complete intermediate time steps, and perform backward inference. The model uses a frozen DINOv2 encoder for static backgrounds and represents dynamic changes with a single 1×768 token, supervising temporal derivatives via JVP. On LIBERO, it predicts 64 frames in just 72 milliseconds with PSNR of 19.46 and LPIPS of 0.134, achieving latency only 1/55th of LDP; this method unifies forward, interpolation, and reverse generation through state change rates.

The team led by Qiu Xipeng at Shanghai Qi Zhi Institute has proposed Embodied Task Agent (ETA) and open-sourced OpenETA, forming a trustworthy closed loop comprising model, tools, runtime, and environmental feedback. The framework separates intelligent decision-making, execution privileges, and environment interaction, enforcing a rule where only one world-changing action is executed at a time, followed by acquiring new observations. It features tool contract validation, execution gating, and immutable trajectory logging. The full version uses GPT-5.6 Luna, succeeding in 56 out of 400 experiments on LIBERO; the lightweight version, equipped with only three tools, achieves 90% Pass@5 on 130 tasks using GPT-5.6 Sol.

Google Cloud has introduced a model routing feature for its API Gateway, currently in public preview. Developers can continue sending OpenAI-compatible requests, which the gateway dynamically routes to models such as Gemini, Claude, or OpenAI OSS-GPT based on configuration—eliminating the need to rewrite separate integration code for different providers. This capability pushes multi-model selection down to the gateway layer, reducing engineering costs associated with model switching and maintaining multiple adapter implementations. Details on general availability timing, pricing, supported regions, and specific routing policies are not disclosed; current availability remains as defined in the public preview documentation.

LendingTree has deployed a multi-agent mortgage assistant on Amazon Bedrock, where a LangGraph supervisor orchestrates tasks and delegates them via MCP to specialized worker nodes. The system routes between Nova Pro and Nova Lite based on task complexity to balance reasoning power and cost efficiency. After deployment, 97% of conversations are completed without human intervention, averaging 6.2 messages per session. Highly engaged users spend over 9 minutes and send more than 10 messages, transitioning from knowledge queries to rate comparisons and pre-approvals. The team attributes improved stability to semantic chunking, conflict resolution in knowledge ranking, full context propagation, and unified checkpointing.