NVIDIA has fully ramped up production of the Groq 3 LPX AI inference accelerator and is offering it as an extension to the Vera Rubin platform for high-speed generative AI and agent tasks. Materials indicate it can increase response speed by 4x, generating tens of thousands of tokens per minute, reducing some agent workflows from hours to minutes. Cloud provider Nebius has become the first customer to commit to deploying the chip. This product signals NVIDIA's expansion beyond GPU training platforms into low-latency inference infrastructure, though pricing, general availability timing, and independent benchmark results have not yet been disclosed.

OpenAI has launched the experimental open standard WebMCP, allowing websites to provide discoverable tool interfaces to AI Agents, enabling Codex to understand and invoke page capabilities without relying entirely on visual operations. Officials stated developers must clearly describe each tool's purpose, applicable scenarios, and permission scope, and continuously refine the toolset through actual usage. WebMCP will be integrated into the ChatGPT desktop app and ChatGPT Sites, automatically using tools when accessing compatible websites. A companion challenge lasting 10 days offers a total prize pool of $35,000, with submissions due by September 3.

Perplexity has released Portable Computer, a locally-prioritized AI Agent platform running on NVIDIA DGX Spark powered by the PPLX 27B model, achieving 82.6% performance on knowledge work benchmarks. The system isolates tasks via OS-level sandboxing, with local steps incurring no token fees. When cloud models are needed, the system first checks context and personal information, then obtains user consent, supporting over 15 different models. Its cloud-upgraded version improves accuracy from 59.6% locally to 73.0%, at a per-run cost of $0.415.

NVIDIA has introduced a shadow engine recovery mechanism to Dynamo, reducing LLM inference failure recovery time from 283 seconds to 7.3 seconds in GLM-5.2 deployment tests. Its GPU Memory Service decouples model weight lifecycle from engine processes, sharing loaded weights via CUDA virtual memory management. Pre-initialized standby engines prepare CUDA contexts, communicators, and computation graphs in advance, delaying only KV cache allocation. After failure, median first-token latency drops from 23,815 milliseconds to 1,311 milliseconds, with decoding speed recovering to 46 tokens per second per user.

OpenAI has launched a $100 Premium seat for ChatGPT Business, primarily targeting small businesses and startups, providing advanced tools previously offered mainly to enterprise clients along with higher usage quotas. The official note states this seat is not subject to the previous 5-hour limit, aiming to support longer-running workflows with more tool usage. Current materials do not disclose billing cycle, specific model quotas, minimum team seat requirements, or full functional differences compared to other ChatGPT business plans. Actual costs should be verified on the official subscription page.

Anthropic has rolled out unified memory for Claude, enabling users to share saved information between Chat and Claude Cowork without needing to configure settings separately across the two interfaces. This update connects long-term context between regular conversations and collaborative workspaces, reducing repetitive input of preferences, background, and project knowledge when switching task modes, and allowing collaboration to build upon prior dialogue history. Existing materials do not specify memory capacity, default enablement status, supported paid plans, enterprise administrator controls, or whether users can isolate workspaces or delete entries individually.

The new version of OpenWorker includes three types of cybersecurity agents designed to scan code vulnerabilities, detect dependency and supply chain injection risks, and check cloud security configurations, helping development teams perform additional security reviews before deployment. Its agent runtime framework is fully open source, allowing security teams to audit data flows and operational logic, reducing the risk of unauthorized code exfiltration. Users can run open-weight models locally to keep sensitive code on-device, or connect to cloud services like ChatGPT or Ox Alpha. Supported platforms, scanning speed, and detection rates have not been disclosed.

Ramp, based on spending data from over 70,000 U.S. enterprise organizations, reports that OpenAI is regaining market momentum relative to Anthropic among enterprise customers. Although Anthropic led in May and July, OpenAI showed faster growth in Q3, while overall enterprise spending on paid AI services continues to rise. This metric is derived from actual expenditure records of Ramp customers, offering a more accurate reflection of procurement behavior than surveys. However, the report does not disclose exact market shares, revenue figures, industry distributions, or sample representativeness, making it better suited for tracking adoption trends rather than estimating total market size.

ChatGPT Work has added secure website login functionality, enabling access to authenticated services without directly exposing user credentials to the model. The company also demonstrated over 20 application scenarios. Its admin plugin now allows viewing 30-day adoption trends, comparing team credit quotas, analyzing spending growth by product and model, and approving quota requests based on current usage and justification—all within a single conversation. These actions leverage existing admin permissions, consolidating investigation, decision-making, and follow-up configuration in one context. Materials do not specify the range of supported websites for secure login or the timeline for broad availability.