AI News from August 12, 2026

NVIDIA outlines how Alibaba’s open-weight Qwen3.8 model runs on GB300 NVL72, with configurable reasoning for large-scale agentic workloads.

Nous Research released NousCoder-14B, an open coding model with a reported LiveCodeBench gain that intensifies competition around AI software development.

OneAdvanced built a UK-sovereign AI platform with 50-plus agents and self-hosted Llama models on AWS, targeting regulated customers with UK data controls.

Anthropic’s Cowork brings Claude Code-style automation to nontechnical users on Mac, while raising new questions about file access, safety, and trust.

OpenAI reports enterprises are moving from AI assistance to agentic execution, with frontier firms using connected tools and workflows far more deeply.

OpenAI’s new Linux desktop app brings ChatGPT, ChatGPT Work, and Codex to major distributions, but omits several computer-use features at launch.

Lovable’s reported valuation has reached $13.3 billion after fresh backing from Tencent and an EU fund, intensifying the race to build AI coding tools.

OpenAI and AWS bring Daybreak Red and Blue to eligible Amazon Bedrock customers, adding governed cybersecurity AI for vulnerability research and defense.

OpenAI is adding $125 Premium seats to ChatGPT Business, offering five times more usage and no five-hour cap as heavier AI workflows raise costs.

Manus is set to resume independent operations after its reported Meta deal unwinds, highlighting geopolitical risk in cross-border AI startup deals.

Anthropic will watermark Claude and other model output, extending the move to older systems as EU transparency rules push AI builders toward traceable content.

Researchers say an API flaw exposed hidden AI reasoning and secrets from public sessions, raising security and privacy risks for AI builders and enterprises.