OpenAI has officially launched GPT-6 Astra, designed for computer operation, software engineering, scientific, and cybersecurity tasks. The model supports a context length of up to 1.05 million tokens, with API pricing at $10 per million input tokens and $50 per million output tokens. It achieves 72.6% on OSWorld 2.0 Offline and 75.2% on DeepSWE v1.1. OpenAI claims the model reaches 99.9% on ARC-AGI-3 under a custom framework, and 62.7% by default, while meeting the Critical cybersecurity threshold. Initial access is granted to select institutions, with rollout to ChatGPT paid tiers, API, and AWS within days.

Meta has released Muse Spark 1.3, optimized for long-horizon agent programming. Internal benchmarks show approximately 20% fewer tool calls and 25% lower token consumption compared to version 1.2. It scores 75.4 on DeepSWE v1.1, surpassing Claude Opus 5 (74.0) and GPT-5.6 Sol (72.7), with a context window of 1 million tokens. The model is gradually available via Muse Code and Meta Model API, maintaining pricing at $1.25 per million input tokens and $4.25 per million output tokens. Weights remain closed, and the highest inference mode is still restricted under security testing.

Security researchers have disclosed two independent exploit chains in the Unitree G1 humanoid robot that enable unauthorized remote code execution (RCE). The first allows an attacker within BLE range to obtain the AES-128 key without pairing, decrypt it via the Unitree cloud API, invoke protected operations, and execute code through a buffer overflow. The second exploits AI service path traversal to upload files to the bashrunner whitelist directory and trigger execution. The report does not specify affected versions, patch status, official response links, or mitigation measures, making it difficult for deployers to assess patch coverage and device risk.

Figure plans to procure up to 100,000 GPUs through Nscale, with a compute commitment exceeding $3.5 billion—termed the largest single investment in computing resources by a robotics company to date, comparable to foundational model vendors. The deal signals a shift among humanoid robot firms from focusing solely on unit hardware to investing heavily in training and inference infrastructure. However, details such as GPU models, delivery timelines, payment structures, or model training plans are undisclosed. This information comes from Physical AI’s summary; no original transaction announcement link is provided, and confirmation awaits official disclosure from both parties.

OpenClaw has launched a native macOS installer and added automatic hardware detection with local model recommendations for Windows machines equipped with NVIDIA GeForce RTX or RTX PRO GPUs with at least 24GB VRAM. It enables running 30B-class models via hosted llama-server and updated llama.cpp. Both platforms now include interfaces for viewing and adjusting proxy permissions. Windows also integrates the open-source Microsoft Execution Containers isolation layer, expected to be fully available in fall. Support for RTX Spark and DGX Station remains in planning.

Cursor co-founder Michael Truell announced the launch of Grok Bot Enterprise Edition, offering two weeks of free access to existing Grok and Cursor enterprise customers. The product has previously been deployed internally within his company, aiming to integrate multiple agents into organizational workflows. However, the announcement provides only qualitative usage feedback, without disclosing official pricing, quotas, supported task scope, data retention policies, or commercial terms post-trial. Availability is currently limited to qualified enterprise customers, and it remains unclear whether general teams or individual accounts will gain access.

LangChain has released an integration tutorial with Nevermined AI, enabling agents to autonomously purchase required resources during task execution without repeated human approval. The solution uses delegated cards with configurable spending limits and supports balance replenishment. All payment records are traceable in LangSmith for auditing agent spending behavior and task provenance. This release is a Cookbook-level reference implementation, not a standalone payment product. The material does not disclose supported currencies, settlement networks, transaction fees, refund mechanisms, or production availability, requiring developers to independently configure permission boundaries.

MongoDB now supports LangChain Deep Agents’ BackendProtocol, allowing agents to continue using interfaces like read, write, glob, and grep while switching the underlying workspace to a persistent backend. With MongoDB Atlas, vector search, full-text search, and hybrid search can be performed within the same storage layer, enabling preservation of cross-session and cross-agent plans, intermediate results, and artifacts. This integration reduces coupling between agent logic and storage systems. However, the material does not specify version requirements, costs, performance benchmarks, or migration tools, necessitating independent validation for production use.