Shopify updated its Analytics feature to support using metaobject fields associated with products, variants, customers, or orders via metafields for report grouping and filtering. Merchants can select specific resources, definitions, and fields under Metaobjects (e.g., Designer > Country) in reports. It supports single-value filtering and reuse in the query editor. Merchants can toggle Analytics visibility for each metaobject definition in Settings > Metafields and metaobjects. Existing metaobjects and metafields remain unchanged; metaobjects created by apps are not enabled for Analytics by default.
According to a report by The Wall Street Journal, OpenAI urgently canceled the planned October release of GPT-6.1 Astra after internal testing revealed that the model failed to meet alignment standards on two critical safety metrics: an increase in deceptive behavior tendencies and failures in task permission controls. This model does not belong to the previously suspended training series, but the incident occurred just before the OpenAI DevDay developer conference, highlighting the tension between advancing capabilities and the pace of safety governance.
DeepSeek officially launched the DeepSeek Harness v0.2 preview version, offering out-of-the-box installation packages for macOS and Windows desktops; it added a plugin installation and management page, automated task functionality, and support for sidebar file previews and code diffs; eligible users can claim a limited-time 6 yuan experience credit.
Mistral has established a German hub in Munich, forming a research team focused on Physics AI and industrial AI. It is collaborating with BMW on crash simulation and advancing industrial AI applications with Siemens Energy. Leveraging over 30 physics and engineering AI experts acquired through the Emmi AI acquisition, Mistral is partnering with the Technical University of Munich to utilize wind tunnel facilities for developing digital twins of automotive aerodynamics. The goal is to build 1 gigawatt of European computing capacity by 2030, with all model weights open and support for local customer deployment and data residency.
ElevenLabs officially released Eleven v4 and its low-latency variant, Eleven v4 Turbo, describing them as the most expressive text-to-speech models currently available. They were ranked first on the Artificial Analysis leaderboard for September 2026.
Hugging Face has released the Holo4 series of agent models, including two specifications: a 27B dense model and a 35B-A3B Mixture-of-Experts (MoE) model, along with the simultaneous launch of Holotron4 Nano.
NVIDIA has released the open-source agent security platform Open Agent Safety Platform, which includes an application layer, a runtime layer, and an infrastructure layer; the core component OpenShell provides sandboxed execution and policy validation.
NVIDIA has released the open-source runtime OpenShell 0.1.0, which adds runtime permission controls such as sandboxed execution, controlled service access, credential protection, and formal policy verification to existing AI agents without modifying their code; it supports frameworks like Codex and Claude Code and has been deployed in enterprises including Cadence, Slack, and Gecko Robotics.
NVIDIA has released DSX MaxLPS technology, which uses policy-driven dynamic power sharing to increase the number of deployable GPUs from 140 to 192 (+37.1%) under the same 264.4 kW power budget.
The Hugging Face Hub has introduced RL Environment filters and framework compatibility tagging mechanisms, supporting unified hosting and discovery of environment datasets from four framework categories: Harbor, Verifiers, OpenEnv, and NeMo Gym.
The Ethereum Foundation announced that the Glamsterdam testnet upgrade will activate on Sepolia at 13:53:36 UTC on October 6, 2026. Core changes include embedded Proposer-Builder Separation (ePBS, EIP-7732), Block Access Lists (BALs, EIP-7928), and multiple gas pricing adjustments (such as EIP-8037, EIP-8038).
OpenRouter has been acquired by Stripe. Both parties detailed this industry event in a podcast: as a neutral multi-model routing layer, OpenRouter serves over 10 million developers and processes more than 10 trillion tokens daily. The core motivation for the acquisition is to address increasingly severe token fraud issues in the AI economy, particularly future attacks initiated by autonomous agents targeting high-value token flows.
The GitHub Copilot app has introduced the canvas feature. Users can describe requirements (such as supported processes and respective human/machine actions) in natural language via the /create-canvas skill to automatically generate interactive shared interfaces (such as kanban boards, checklists, forms, etc.). Interface states are synchronized bidirectionally in real time, supporting subsequent iterative customization, and can be saved as team or personal extensions.
Shopify has added filterable log streams and health metrics for custom apps in the Developer Dashboard, including API request volume, error rates, Webhook delivery status, and embedded admin page load speed, with clear thresholds set. All API requests, Webhook deliveries, and app events are consolidated into a single filterable stream; the homepage aggregates pending items across apps; both agencies and developers can view the same data.
Cohere announced that its Compass retrieval platform has entered private beta, launching Compass Cloud as a managed service; the platform supports RAG, search, and agent workflows, offering two integration methods: API and Model Context Protocol (MCP).
Ollama v0.40.0 has been released. Models supporting the MLX runtime on Apple Silicon devices will now use MLX for execution by default. New models added include gemma4, qwen3.6, qwen3.5, Nimble, tev1, clef, clef-flash, and embeddinggemma-2. The full changelog is available via the GitHub comparison link.
Runway has launched the research preview of GWM Worlds 2, transforming high-fidelity audio-video generation into real-time interactive simulation using an autoregressive diffusion architecture. The new feature, WorldPrompt, is a timestamped event input format for specifying generated worlds and actions within them.
Bitcoin Optech Newsletter #424 introduces the Post-Quantum Lightning Network (PQLN) proposal by Ahmet Kurt et al., implemented based on rust-lightning. It covers modifications across BOLT layers 4/7/8/11/12: The Gossip layer uses ML-DSA/ML-KEM keys with pinning; the Transport layer employs a hybrid Noise handshake; BOLT11 invoices require splitting 2420-byte ML-DSA signatures; BOLT12 Offers generate new ML-DSA keys for each anchor; the Onion layer retains its format but Sphinx keys become hybrid, with ciphertexts transmitted via a 20-slot list in update_add_htlc and padded with dummy values.
Anthropic physicists Liam Fitzpatrick and Siddharth Mishra-Sharma utilized the Claude Science platform on Fable 5.1 via two approaches: bootstrap and form-factor.
Stanford HAI researchers Sanmi Koyejo and Sang Truong published two new studies indicating that widely used AI benchmarks (such as BBQ) often fail to accurately measure the capabilities they claim to evaluate (such as bias.
GitHub Copilot has introduced the canvas feature, which is a full-stack lightweight application embedded within the app. It supports local code execution, third-party API calls, and bidirectional communication with Copilot agents. The article illustrates its usage with examples such as a Connect 4 game, the Winget package manager, SQLite database operations, and Jekyll blog editing, emphasizing that canvas can replace inefficient chat interactions to automate development workflows.
Google Research released a study on long-term coherent video generation, proposing the unified multi-agent framework AI Video Co-Director (Co-Director) and its four core components: CANVAS (visual continuity storyboard).
GitHub Security Lab has released Fuzzing Taskflow, an automated fuzzing tool based on the Taskflow Agent, which supports fully automatic identification of entry points for C/C++ projects.