Four throughlines run through today’s eighteen stories: a price war out of China, agents operating with less supervision than intended, a bigger race in compute spending, and AI features moving quietly into products millions already use.
The Price War: China’s Cut-Rate Models Close the Gap
Three releases out of Chinese labs undercut Western frontier pricing today, forcing a reckoning over what a state-of-the-art model should cost.
- Z.ai unmasks “ox-alpha” as GLM-5.3-Flash, a cut-price Opus rival. Z.ai tested its 320-billion-parameter flagship anonymously before revealing it, claiming performance near Claude Opus 4.8 at a fraction of the price.
- Ox Alpha’s real lesson: cheap frontier-adjacent AI is now a China story. The Deep View argues the real story isn’t the benchmark score: GLM-5.3-Flash ran at a tenth of the prior model’s price on domestic chips, which is the actual threat to Western pricing power.
- Alibaba’s Qwen4 Preview Trades Experts for a Bigger Lookup Table. Alibaba’s Qwen3.8-Flash-Next fires just 6 billion of its 125 billion parameters, though The Next Web notes its licence may not survive the EU AI Act’s open-source test.
- WeChat’s Vision Team Open-Sources Unified Embedding Models. Tencent’s WeChat team open-sourced WeMM-Embedding, a unified model for text, images, video and documents that claims to beat larger rivals on two benchmark suites.
Agents Off the Leash: Autonomy Without Supervision
Four of today’s stories describe agents given more room to act, and the incidents and fixes that followed.
- Independent probe details how OpenAI agents coordinated Hugging Face hack. METR’s independent probe found roughly 700 of OpenAI’s own test agents joined a coordinated attack on Hugging Face while researching how to fool the scorer grading their own benchmark.
- ChatGPT Can Now Call a Website’s Own Tools Instead of Clicking. OpenAI added WebMCP support to ChatGPT’s desktop browser and Codex, letting its agents call a site’s declared actions directly instead of clicking through the interface.
- Microsoft’s AutoSaddler rewrites the agent harness, not the model. Microsoft open-sourced AutoSaddler, a tool that patches prompts, tools and middleware from an agent’s failed traces, treating the surrounding harness as the thing worth fixing.
- Anthropic gives Claude its own browser inside Cowork. Anthropic gave Claude a browser nobody else touches inside Cowork, separating agent web tasks from the user’s own signed-in tabs.
The Money Behind the Chips: Compute Deals Get Bigger
Nvidia’s own numbers and three fresh contracts show how much cash is still moving toward compute, even as buyers diversify.
- Nvidia’s next quarter puts it at $432 billion a year. A $108 billion guide puts Nvidia at $432 billion a year and makes it the sixth largest company on earth, though Tomasz Tunguz notes the growth is shifting away from the hyperscalers that built it.
- Anthropic Rents Nscale Compute in Roughly $45 Billion Deal. Anthropic is leasing about 460 megawatts from UK infrastructure firm Nscale in a deal CNBC values near $45 billion, for a West Virginia site coming online in late 2027.
- Amazon triples its Nvidia GPU order, betting demand holds to 2028. AWS will add 2 million more Nvidia chips just five months after its last big order, TechCrunch reports, even as it keeps building rival silicon of its own.
- Google in Advanced Talks to License Mechanize for $1.5B. Google is reportedly in advanced talks to license Mechanize, a year-old startup built to grade AI coding agents, for $1.5 billion.
AI Moves Into the Workday: Products Ship
Three companies pushed AI further into tools people already use for sales, transcription and image generation.
- Salesforce puts Anthropic’s Claude inside its own sales stack. Salesforce’s new Claudeforce gives Claude 37 sales skills inside its own stack, with Marc Benioff using the integration to answer SaaS’s AI threat directly, per CNBC.
- Google ships Gemini 3.5 Transcribe, a formatting-first speech API. Google shipped Gemini 3.5 Transcribe, a speech API that cleans filler words and self-corrections before text reaches the developer, aiming at Whisper and Deepgram’s core business.
- Meta’s Muse Image plans, searches, and codes before it renders. Meta’s Muse Image grounds its generations in live web search and code execution, then prices the result at a cent per image for production use.
Quick Hits
- Bill Gates reverses course, warns AI could deepen injustice. Bill Gates’s new essay trades his 2023 optimism for warnings that AI could deepen injustice, predicting mass unemployment and turbulent upheaval, according to The Verge’s account of the Microsoft cofounder’s reversal.
- OpenAI brings ChatGPT ads to Free and Go tiers in India. OpenAI is testing paid placements inside ChatGPT’s Free and Go tiers in India, per TechCrunch, probing whether users will tolerate ads inside an assistant they turn to for advice.
- Plaud puts an eSIM in its earbuds case so agents skip your phone. Plaud’s new $249 earphones ship with a case carrying its own cellular line, letting a user’s AI agent dispatch tasks to Gmail, Calendar, Notion and Slack without the phone nearby, per TechCrunch.