Nvidia says AI agent failures should be viewed as a system problem, with causes possibly in the harness or runtime rather than the model. The company is developing the OpenShell runtime and launched the SAFE initiative with about 140 companies to share agent failure data.
Microsoft added a side panel with a built-in browser to Copilot: source links open directly in the chat window without interrupting the dialogue. The behavior is similar to built-in browsers in Teams and Outlook.
Google Labs expanded access to the experimental AI agent CC: the tool is now available to families and households for managing schedules and chores. CC previously launched last year in a limited mode.
DeepSeek AI open-sourced DeepSeek Harness (dsh), an agent environment where session storage, sandbox, context compression and the model itself are plugins. The MIT-licensed project is in developer preview, runs on Cordis with profiles and patches reloadable on the fly.
Google DeepMind is preparing the Gemini 4 Pro model for content creation by October 2026: SVG generation, 3D modeling and interactive simulations. Early tests showed complex vector graphics and detailed 3D projects; the model will compete with GPT-6 Astra, Fable 5.1 and Opus 5.2.
WIRED tested Meta's AI agent Muse: the app was downloaded over 900,000 times in the first week. The agent can order food and find items on Facebook Marketplace, but pushes email and bank account connections, and user data is used for model training by default.
AI developer Prinz used OpenAI's GPT-6 Astra model to decrypt a 108-year-old German radio message from World War I. The key was the word TRUPPENVERSCHIEBUNG, and the result matched military logs: British cruiser HMS Canterbury arrived in Sevastopol on November 24, 1918.
Google released Agent Development Kit for Kotlin 1.0, a production framework for building AI agents on Kotlin, Android, and JVM. The version reaches parity with ADK for Python and Java and adds support for local and hybrid AI on Android.
At HUAWEI CONNECT 2026 in Shanghai, Huawei computing strategy director Zhu Zhaosheng said Ascend has passed a key ecosystem inflection point and presented a five-dimension agentic computing path for SuperPoD and SuperCluster. The CANN community exceeds 5,200 monthly active users, external contributors outnumber Huawei employees, and partners gain access to resources at 10,000 NPU scale.
StepFun introduced Step 5 Preview, a sparse MoE model with ~600 billion total parameters (27 billion active per token), 1 million token context, and support for text and images. The API is already available at $1 per 1 million input tokens and $2.7 per 1 million output tokens, with weights promised to open on October 15.
Qwen introduced the Qwen3.8-LiveTranslate simultaneous translation model with Interleave architecture, cutting average LAAL latency from 2.8 to 2.3 seconds. It understands 60 languages, voices 29, supports speaker diarization and bilingual output, and is available via Alibaba Cloud Model Studio and QwenCloud.
The anonymous stealth/union-alpha model on OpenRouter, launched September 16, turned out to be the blended Pareto model from unbiased.ai. The free period was closed early due to a load of 1B tokens per minute; the current Pareto 26.9 costs $2.50 per 1M input and $7.50 per 1M output tokens.
OpenAI released the legal service Astra for Law on GPT-6 Astra with a search index of 230 million URLs, including 99.9% of published US case law. Legal search accuracy rose 40%, but correctness on the Vals AI Legal Research Bench was only 54.0%. Early access is available via Trusted Access in ChatGPT and Codex.
OpenAI introduced GPT-5.6 Sol with time to first token under 100 ms and throughput of 180 tokens/s. Pricing is $4 per 1 million input tokens and $20 per 1 million output tokens. The FlashDecode architecture keeps a warm cache of the model's initial layers refreshed every 60 seconds.
The McKinsey Technology Trends Outlook 2026 report found that productivity declined in 30% of companies after adopting agentic AI tools. A 180% rise in coding activity yielded only a 30% increase in releases, and 46% of developers do not trust AI accuracy.
The HyQuant method quantizes most query, key, and value tensors to low precision, keeping only critical tokens and a local sliding window in full precision. This yields 1.32–3.58x speedup of the decoding kernel and 1.04–1.17x end-to-end decoding with less than 1% accuracy loss.
Nvidia VP Adel El Hallak said he uses DGX Spark at home for overnight AI tasks. According to him, local models on own hardware provide privacy and no subscription, and this fall RTX Spark laptops with the same GB10 chip and 128 GB memory will launch from Dell, HP, Lenovo, Asus, MSI, and Acer.
Chinese startup Spirit AI expects a breakthrough in embodied AI by mid-2027, enabling robots to perform tasks via verbal commands. The company is valued at 20 billion yuan ($2.9 billion), and its Moz1 robots already work on CATL and JD.com lines.
Alibaba introduced the open MoE model Qwen3.8-Flash-Next: 125 billion parameters with 6 billion active per token and a 51 billion N-gram embedding table. Context is 262,144 tokens extendable to 1 million, claimed score of 62.5 on SWE-bench Pro, and training cost about 1/9 of Qwen3.7-Plus.
Zhongguancun Academy and Zhongguancun AI Institute released ZGCM-1-7B, a dense 7.39B-parameter model with open weights, data, and training code under MIT license. It uses hybrid attention (27 gated sliding-window layers and 5 global), 256K token context, and thinking and direct answer modes; trained on ~4.19 trillion tokens.
At HUAWEI CONNECT 2026 in Shanghai, Huawei Cloud introduced its agent stack: AgentArts platform, open-source openJiuwen, Agentic Model as a Service, and Industry AI Foundry. AICS will launch in China on September 30, overseas on November 30, and AgentArts overseas on December 30.
Huawei launched the 100 NPU-Hour program, giving developers access to NPU-based computing power. The initiative opens access to 10,000 NPUs for training and running AI models.
AI lab Martian introduced Code Review Bench, an open benchmark for AI code review tools. On 16,017 real GitHub pull requests, Cubic Dev AI leads with F1 65.7%, followed by GitHub Copilot 63.9%, Claude 62.5%, CodeRabbit 60.8%, CodeAnt AI 50.0%. Methodology and code are published under the MIT license.
A BBC survey showed that many employees at OpenAI, Meta, and DeepMind are skeptical of claims that uncontrolled AI will destroy humanity. Former Anthropic employee Jacob Coxon said AI agents could create bioweapons but gave no details. Anthropic brought in external evaluators from Faculty, owned by Accenture.
Open personal AI agent OpenClaw released version 2026.9.5 with 4179 pull requests from 502 contributors. The main innovation is atomic updates: the new version is tested on a private copy of the configuration, and rollback occurs on failure. Also added plugin hot reload, session sharing, and expanded GPT Live for meetings.
Binance introduced Agent OS, infrastructure for trading via AI agents through MCP. Agents operate in a separate subaccount with set permissions but cannot withdraw funds; trading is limited to the amount the user deposits.
A user noticed that Meta's AI assistant Muse knew the contents of his messages, though he had not granted access. Meta explained that Muse does not monitor Mac notifications and syncs Messages data only after explicit permission, and admitted the assistant gave an incorrect description of its own functions.
An XDA author tested a leaked DLSS 5 build with Neural Rendering on The Witcher 3, Batman: Arkham Knight, and Metro Exodus. Landscapes and lighting effects became more photorealistic, but familiar characters' faces turned into 'strangers' — likely because the model was trained on ordinary human faces.
The MIT-licensed Cua project (trycua/cua) released a platform for computer-use agents: background drivers for macOS, Windows and Linux, Cua Fleets cloud sandboxes, compact CUA-S1 models for fast UI decisions, the Lume VM manager for Apple Silicon and the Cua Bench benchmark.
According to an NYT investigation, DraftKings built an ML model in 2023 that scored customers by expected losses to target bonuses. A problem gambling detection tool was never completed. The company says AI personalization of promos improved sportsbook margin by 13% in 2025.
