Microsoft launches new in-house AI models it says cut costs up to 89% versus OpenAI
Microsoft released two in-house AI models, MAI-Image-2.5-Pro and MAI-Voice-2-Flash, claiming production costs 89% lower than OpenAI's comparable models. These models now power Bing, PowerPoint, OneDrive, Dynamics 365, Excel, GitHub Copilot, and Azure for millions of users.
Andrew Ng Just Released OpenWorker: An Open-Source, Local-First Desktop AI Coworker That Returns Finished Deliverables Instead of Chat
A desktop AI agent called OpenWorker returns finished work deliverables instead of chat responses, running locally through a Python server within a Tauri shell. The system supports 30 tool-calling models plus local Ollama, and requires approval for all write operations, shell commands, and external actions through a risk engine.
Multi-turn attacks broke AI models 88% of the time — single-turn testing missed it, Cisco AI security lead warns at VB Transform 2026
Multi-turn attacks broke through 15 flagship AI models 88.3% of the time when attackers adapted across conversations, while single-turn testing missed these vulnerabilities. Only 32% of enterprises surveyed give every agent its own scoped, managed identity, and 30% isolate high-risk agents in sandboxes.
The agent security gap: 54% of enterprises have already had an AI agent incident, and most still let agents share credentials
More than half of 107 surveyed enterprises experienced confirmed AI agent security incidents or near-misses, with only about one-third assigning each agent its own scoped identity. Most agents still share credentials, only 30 percent isolate high-risk agents, and enterprises rely on security tools borrowed from model providers rather than purpose-built agent controls.
The AI context gap: Enterprise AI organizations have a trust problem, not a retrieval problem — and most are still building the fix
A majority of 57 percent of enterprises experienced their AI agents producing confident but incorrect answers traced to missing or inconsistent context over the past six months. Provider-native retrieval has overtaken dedicated vector databases as the default context source, while most enterprises are still building governed semantic layers to address trust issues in their retrieval-augmented generation systems.
Nvidia bets physical AI can solve healthcare robotics’ data problem
Nvidia launched a Medical Physics Simulation framework that trains healthcare robots through physical interaction rather than code alone. The framework treats robots as physical AI systems requiring embodied experience to learn how the world behaves through contact and force.
Best Open Speech Recognition (ASR) Models in 2026: WER, Languages, Latency, and License Compared
Multiple open speech recognition models now score within one word error rate point of each other on the Hugging Face leaderboard, ending Whisper's dominance. Cohere Transcribe, IBM Granite Speech 4.1, ARK-ASR, and MOSS-Transcribe compete alongside 12 other open-weight models measured by word error rate, language coverage, streaming latency, and license terms.
Meet Gigatoken: A Rust BPE Tokenizer that Encodes Text at 24.53 GB/s, up to 989x Faster than HuggingFace Tokenizers
A Rust BPE tokenizer called Gigatoken encodes text at 24.53 GB/s on a 144-core AMD EPYC processor. It achieves 989x faster performance than HuggingFace Tokenizers through a hand-written SWAR pretokenizer and pretoken caching.
Anthropic Releases Claude Security Plugin for Claude Code in Beta: A Multi-Agent Vulnerability Scanner That Runs in Your Terminal
A multi-agent vulnerability scanner that operates within Claude Code sessions has entered beta testing. The plugin scans repositories for vulnerabilities and converts selected findings into patch files for user review and application directly from the terminal.
NVIDIA AI Supercomputer Comes Online at Naval Postgraduate School
A DGX GB300 AI supercomputer has been activated at the Naval Postgraduate School in Monterey, California for use by students, researchers, and faculty. The system represents one of the world's most powerful AI platforms now operational at the U.S. military's flagship graduate university.
Cursor Releases Cursor Router: A Request-Level Classifier Delivering Frontier Coding Quality at 30–50% Lower Cost
Cursor Router classifies coding requests by query, context, complexity, and domain to route them to appropriate models, achieving 30–50% cost savings compared to Opus 4.8 rates. The system is now available for Teams and Enterprise plans and delivered frontier-quality output in early testing with three enterprise accounts.
AI agents aren't confidently wrong because of bad context — they're wrong because of bad data engineering
AI agents confidently provide incorrect answers when underlying data stores become outdated, such as when pricing changes or product versions update, because retrieval systems score relevance rather than correctness. Standard retrieval pipelines lack mechanisms to detect whether served data remains accurate, allowing stale documents and incomplete records to pass through undetected.
SenseTime’s Galaxy Project targets domestic AI chip scale-up
SenseTime launched the Galaxy Project with nearly 20 partners to scale domestic AI chip infrastructure in China. The initiative aims to create a closed loop connecting chip-level technology and ecosystem development.
Cisco Foundation AI Releases Antares: 350M and 1B Open-Weight Models That Localize Known Vulnerabilities Inside Real Codebases
Cisco Foundation AI released Antares, open-weight models of 350M and 1B parameters trained to locate known vulnerabilities in codebases. The 1B model achieved 0.209 File F1 on the Vulnerability Localization Benchmark, outperforming larger models, while a 500-task evaluation costs under one dollar on a single H100.
Poolside drops Laguna S 2.1, an open-weight coding model that beats rivals 10x its size
Poolside released Laguna S 2.1, a 118-billion-parameter coding model that activates only 8 billion parameters per token and achieves 70.2% on Terminal-Bench 2.1, outperforming models several times larger. The model supports a 1 million token context window and is available on Hugging Face under an open license.
Stop adding more GPUs: Weka's new storage platform reduces load by caching 100% of an AI model's pre-calculated tokens
Weka's NeuralMesh 6 platform caches 100% of an AI model's pre-calculated tokens using cheaper NAND flash storage instead of expensive GPU memory. The system aggregates flash storage to function as GPU memory at a fraction of the cost, reducing the need for additional GPU resources in production AI workloads.
Google's Gemini 3.6 Flash model cuts AI agent token costs by up to 65% on long horizon engineering tasks —and 3.5 Pro is on the way
Google released Gemini 3.6 Flash, which reduces AI agent token costs by up to 65 percent on long horizon engineering tasks and is priced at $1.50 per million input tokens and $7.50 per million output tokens. The company also released Gemini 3.5 Flash-Lite at $0.30/$2.50 per million tokens, though the prior generation Gemini 3.1 Flash-Lite remains cheaper at $0.25/$1.50 per million tokens.
Evals are the new PRD, Expedia’s AI chief tells VB Transform 2026
Evaluation frameworks are replacing traditional product requirement documents by encoding desired product behavior through tests including red teaming and security checks before coding begins. Sixty-six percent of surveyed enterprises allow production deployment without human review or plan to within 12 months, yet only 5% fully trust their automated evaluations.
Google Releases Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber: A Cheaper, More Token-Efficient Flash Tier Built for Agentic Workloads
Google released Gemini 3.6 Flash with 17% fewer output tokens and output pricing of $7.50 per 1M tokens on July 21, 2026. Flash-Lite operates at 350 tokens per second while gated Flash Cyber powers CodeMender for vulnerability finding.
Moving beyond passive RAG: How to implement active memory reconstruction for AI agents
Active memory reconstruction improves LLM reasoning by filtering noise from retrieval-augmented generation systems and reducing token consumption. MRAgent implements this approach to enable AI agents to selectively retrieve and reconstruct relevant memories during task execution.
The AI Slot Machine Effect: Why Generative Feeds Disrupt Deep Work And How to Reclaim Focus
Generative AI tools cause users to spend extended time refining prompts repeatedly, diverting them from their original tasks. Knowledge workers in 2026 increasingly experience this distraction pattern when using these tools for quick productivity boosts.
The Download: Chinese AI divides the White House, and a record copyright payout
Chinese AI models have become a source of conflict among President Trump's advisers, with current and former officials publicly criticizing each other over how to handle the technology. The disagreement reflects division within the White House over China's artificial intelligence development and policy response.
Meta Open-Sources Astryx: An Agent-Ready React Design System With 150+ Accessible Components, Seven Themes, and a CLI
Meta open-sourced Astryx, a React design system containing 150+ accessible components and seven themes that was used internally across 13,000+ apps for eight years. The system requires React 19 or later and includes dark mode, templates, an agent-ready CLI, and is released under the MIT license.
Chinese open-weight models are cheap. Washington is deciding what that costs.
Moonshot AI released Kimi K3 on July 16 as the largest open-weight model yet released, reopening policy debates in Washington. Enterprises now question whether using Chinese open-weight models will remain feasible given potential regulatory changes.
NVIDIA Releases Cosmos 3 Edge: A 4B-Parameter Open World Model That Reasons and Generates Robot Actions On-Device
A 4-billion-parameter open world model called Cosmos 3 Edge enables robots and vision AI agents to understand surroundings and generate actions locally without cloud dependency. The Cosmos 3 family includes three versions: the 4B Edge model, a 16B Nano variant, and a 64B Super model released on May 31, 2026.
Atlassian: Research shows organizations should approach AI at the team level, not the individual level, to achieve true ROI
Eighty-nine percent of executives report individuals are speeding up with AI adoption, yet only 6% can identify specific ROI examples. Atlassian's research of 12,000 knowledge workers and 200 Fortune 1000 executives found that approximately 14% of teams translated AI usage into measurable value.
Writer's AI harness cuts token spend nearly 40% — without sacrificing accuracy
Token consumption for AI applications dropped nearly 40% through optimizations to the orchestration layer surrounding foundation models, without reducing accuracy. Engineering teams achieved these reductions by systematically refining the AI harness components that wrap the model, lowering cost-per-successful-task by up to 61%.
Alibaba’s Tongyi Lab Releases Qwen-Audio-3.0-TTS, a Hosted Text-to-Speech Model in Flash and Plus Tiers Across 16 Languages
Alibaba's Tongyi Lab released Qwen-Audio-3.0-TTS, a text-to-speech model available in Flash and Plus variants supporting 16 languages. The model is hosted through Alibaba Cloud Model Studio rather than offered as downloadable weights, with Flash optimized for real-time interaction and Plus for high-quality audio generation.
A single AI agent conversation can look perfect and still be broken, leaders from LangChain, Conviva and CoreWeave said at VB Transform 2026
Enterprises are shifting away from scoring individual AI agent conversations toward comparing cohorts of users against baselines, because single conversations can appear flawless yet indicate broken products. The industry faces tension between scalable but ungrounded automated judging through LLMs or agents and human review that cannot scale.