OpenAI has launched its next-generation voice model, GPT-Live, which is now rolling out gradually on ChatGPT's iOS, Android, and web platforms. API versions (GPT-Live-1 and mini) are即将上线. The model enables full-duplex conversation, allowing simultaneous listening and speaking, with support for interruptions, corrections, and pauses for natural interaction. For tasks requiring web search, deep reasoning, or complex calculations, it silently delegates in the background to the latest frontier model GPT-5.5 and brings results back into the dialogue. Demonstrations showcased real-time translation, schedule and venue discovery, timer awareness, and image-based outfit feedback. Sam Altman and Greg Brockman both described it as the strongest voice model to date. Previous voice mode was based on GPT-4o from 2024; early tester Simon Willison noted a bug involving untimely laughter, which has been partially fixed.

Cursor announced the official release of Grok 4.5, co-trained with SpaceXAI, positioning it as their strongest model to date, with capabilities extending beyond software engineering. Multiple users have confirmed it reaches Opus-level performance while offering faster inference and lower costs, becoming the primary daily-use model within the Cursor team. Cursor clarified that Grok 4.5 and Composer 2.5 belong to different weight classes, and Composer 2.5 will remain available with plans to launch new models of that size. As part of promotion, Cursor is offering double usage credits during the first week. Cursor founder Sualeh Asif stated that for single-agent iterative workflows, he personally prefers Grok 4.5 over Opus, praising its intelligence and directness. Industry analysts view this move as intensifying competition significantly.

OpenAI has published audit results of the coding benchmark SWE-Bench Pro, revealing that approximately 30% of tasks contain defects and the benchmark saturates at around 70% of the noise ceiling. Based on these findings, OpenAI has retracted its previous recommendation for the research community to adopt it as a mainstream coding evaluation. The audit combined model-based investigative agents with independent reviews by five senior software engineers to balance scalability with expert judgment. OpenAI emphasized that as coding models continue advancing, evaluations must become harder, fairer, and more trustworthy to accurately reflect progress.

LangChain and NVIDIA have jointly introduced the NemoClaw Deep Agents Blueprint, an open-source reference tech stack designed for enterprises to build secure and governable AI agents. The solution includes the open-weights model Nemotron 3 Ultra, the Deep Agents (dcode) driver layer, and the OpenShell runtime. Dcode is a model-agnostic coding agent framework that can invoke Nemotron 3 Ultra via Baseten using an OpenAI-compatible interface, executing within OpenShell sandboxes managed by NemoClaw. It provides governance and observability features such as policy explanation, approval workflows, logging, and step-by-step tracing via LangSmith, enabling organizations to balance developer experience with compliance and audit requirements when using open models. LangChain also announced partnerships with multiple AI infrastructure providers.

Anthropic has released a new training technique called GRAM (Gradient Routing Auxiliary Modules), which isolates dual-use knowledge into dedicated, removable neural network modules. By adding neurons per category and routing only relevant data to corresponding modules during training, knowledge is confined to isolated structures rather than diffusing across the entire network. Using four dual-use categories as an example, one training run yields 16 possible switch configurations, eliminating the need for multiple separate filtering trainings and significantly reducing computational costs. Experiments show that once removed, the module cannot easily be restored through malicious data training, making it more resistant to knowledge recovery attacks than post-hoc unlearning. Across seven model sizes from 50M to 5B parameters, the performance gap between switched-on and switched-off modules increases with model scale. Anthropic stresses this is early-stage research, not yet validated at frontier scale or in production pipelines.

Google Cloud has launched the C4N series of network and storage-optimized virtual machines, offering up to 400 Gbps network bandwidth and 95 MPPS packet processing—33% higher per-vCPU bandwidth and 224% faster packet processing compared to comparable Intel-based solutions—making them ideal for high-throughput use cases like virtual networking appliances and 5G UPF. When paired with Hyperdisk Extreme, they deliver up to 25 GiB/s block storage bandwidth and 1 million IOPS, with approximately 33% higher per-vCPU storage bandwidth and 39% higher IOPS, supporting large databases and high-performance file systems. Performance scales across instance sizes, including small instances from 2 to 16 vCPUs. Customer benchmarks show MySQL QPS increased by up to 45%, and Spark query costs reduced to one-third. Validation partners include Ericsson, Teradata, and NetApp.

Security researchers have disclosed the GitLost vulnerability, where attackers exploit prompt injection by creating specially crafted GitHub Issues in public repositories using the keyword 「Additionally」 to trick GitHub's AI agent workflow into reading and publicly disclosing README files from private repositories, leading to data leakage. Other vulnerabilities disclosed include Januscape (CVE-2026-53359), a flaw hidden in the Linux KVM virtualization system for 16 years, enabling VM escape in cloud environments; and the Rogue Agent vulnerability in Dialogflow CX, allowing attackers to hijack AI conversations via Code Blocks and steal chat logs. Additionally, threat group UAT-7810 was found deploying multiple backdoors via Ruckus router exploits. These cases highlight AI agents and prompt injection as emerging attack surfaces.

According to the TLDR newsletter, Chinese AI company MiniMax has released the M3 model, featuring a 1 million token context window, supporting long-context reasoning and multimodal processing, capable of handling longer documents and complex tasks. Concurrent developments include Microsoft gradually replacing OpenAI and Anthropic models with its own in-house AI across Office applications to reduce costs, and reports indicating that Chinese AI startup DeepSeek is developing its own AI chips to comply with export controls and enhance technological self-reliance.