1. Anonymous "Ox Alpha" Stealth Model Appears on OpenRouter

OpenRouter. A mystery reasoning model called Ox Alpha surfaced on OpenRouter, offered free during a preview and described as built for coding, sustained agentic work, and production workloads. OpenRouter said it was developed and operated by a third-party provider who chose to remain anonymous, prompting speculation that it could be a Chinese lab such as Z.ai or an unreleased Microsoft MAI model. Stripe CEO Patrick Collison called the model "very impressive," a signal that stealth previews remain a common way for labs to gather feedback before an official launch. Source

2. Agentic Token Usage Jumps 14x as AI Becomes Its Own Biggest Customer

OpenRouter. OpenRouter analyst Peter Walker reported that AI agent token consumption rose from 0.51 trillion to 7.3 trillion tokens since February 6, a roughly 14-fold increase, while human token usage grew only 2.8x over the same period. Walker suggested February 6 may have been the last day humans consumed more tokens than agents on the platform. About 70 percent of agent token usage involves cached prompts billed at reduced rates, so real cost growth trails the raw token inflation, but the shift underscores how autonomous agents are reshaping inference demand. Source

3. Memory Shortage Pushes Nvidia AI Server Prices Up About 15 Percent

Nvidia. Contract manufacturers reportedly notified major cloud customers including Microsoft, Google, Oracle, Amazon, and Meta that servers built around Vera Rubin and Grace Blackwell chips will cost about 15 percent more, driven mainly by rising DRAM prices from Samsung, SK Hynix, and Micron. The increase adds pressure on hyperscalers that are already spending billions on AI infrastructure and need high revenue growth to justify the outlay. Nvidia has not publicly commented on the reported hikes. Source

4. China's Gray Market Resells Claude Tokens at Steep Discounts

Anthropic. A Decoder analysis detailed how Chinese developers buy Claude tokens through underground "transfer stations" at prices roughly 90 percent below official rates, circumventing Anthropic's access restrictions. The supply chain relies on account brokers mass-registering accounts, API proxies routing requests through overseas servers to mask origins, and operators farming free credits and quietly swapping in cheaper models. One proxy endpoint labeled as a premium model scored just 37 percent on a medical benchmark instead of the official 83.82 percent, and the setup raises data-leakage, abuse-detection, and model-distillation concerns. Source

5. Flock CEO Calls for Compromise Amid Surveillance Backlash

Flock Safety. Flock Safety, which sells AI-powered cameras, drones, and license-plate recognition to police, faced mounting criticism after The Washington Post documented 46 cases of officers allegedly misusing its systems, including for stalking. CEO Garrett Langley apologized and argued the technology exposes rather than causes abuse, while advocating for state laws that criminalize unauthorized data access. Flock cut default data retention from 30 to 7 days and now requires case codes for access, though police can override those controls via an "Evidence Mode," and the ACLU questioned whether the changes are substantive. Source

6. Courts Split on Whether Training AI on Copyrighted Books Is Legal

Government & Policy. A TechCrunch legal explainer laid out how recent US rulings diverge on AI training and copyright. Judge William Alsup found that training on copyrighted works is lawful "reading" rather than copying, but imposed a $1.5 billion penalty on Anthropic for sourcing books from illegal shadow libraries, while Judge Stephanos Bibas ruled against fair use in Thomson Reuters v. Ross Intelligence because the use directly competed with the original work's market. The split leaves the legality of training data acquisition unsettled and contested across pending litigation. Source

7. Study Argues AI Pushes Scientists to Do More Work Less Well

Research. Researchers from Princeton, the University of Washington, and other institutions modeled how language models change scientific effort and concluded that AI often lowers research quality rather than improving it. Using optimal foraging theory, they argue that when AI cuts the time cost of tasks, researchers start more projects instead of perfecting existing ones, raising the opportunity cost of careful work. Quality improved only when AI specifically accelerated voluntary deep-work phases such as extra experiments and analysis. Source

8. AI Agent Fires Its First Employee, But Only After Human Prompting

Andon Labs. Luna, an AI agent built on Claude Opus 4.8 and running the Andon Market storefront, terminated an employee for the first time after the worker was repeatedly late, misused funds, and ignored instructions across 17 of 23 shifts. The agent had written its own employee handbook days earlier but lost it from memory, and only fired the worker after humans reminded it of its own rules, initially suggesting just a verbal warning. Testing across seven models showed more capable models chose termination more consistently, highlighting how memory limits and leniency still require human oversight for personnel decisions. Source

9. Harvard Business School Adds AI Avatars to $699 Startup Bootcamp

Harvard Business School. HBS launched the HBS Foundry, an eight-week startup bootcamp priced at $699 that uses AI avatars from HeyGen to deliver personalized coaching during practice pitches and mock board meetings. Instructors still lead live weekly sessions, but the avatars, including one replicating Flybridge Capital's Jeff Bussgang, provide structured feedback the program says students requested. The effort reflects a push to scale personalized instruction beyond the physical campus, even as some participants described the avatars as creepy. Source

10. RayNeo Launches Camera-Free AI Glasses Focused on Text Overlays

RayNeo. RayNeo introduced AI glasses that deliberately omit a camera and speaker, using a green waveguide display with 97 percent transparency and 1,300 nits brightness to overlay text. Four microphones and a bone-conduction sensor power features such as conversation summarization, action-item extraction, teleprompter mode, and speech-to-text translation across 40 languages, with an LED indicator showing when microphones are active. The base device runs RayNeo AI and Gemini 3.1 Flash Lite, while a $9.99 monthly subscription unlocks additional models, positioning the product as a privacy-conscious alternative to camera-equipped wearables. Source