Anthropic reveals bioweapon filters were off for 133M contractor chats

Anthropic ran 133M contractor chats with bioweapon filters disabled Anthropic's Risk Report says blocking classifiers were off on human-feedback traffic from May 2025 to April 2026, and an internal-use flag also disabled logging. Roughly 50,000 contractors generated about 133 million exchanges vetted by vendors that often lacked strong screening; Anthropic found no clearly concerning misuse but raised its catastrophic-misalignment risk from "very low" to "low." 1

Anthropic Q2 revenue tops $11.5B as IPO uses 2028 forecasts Preliminary Q2 revenue of more than $11.5B compares with $787M a year earlier and $4.73B in Q1, with positive adjusted operating income. Bloomberg reports bankers are valuing Anthropic on forward revenue multiples, with the company projecting roughly $190–200B in 2028 revenue ahead of a potential blockbuster IPO. 23

Jeff Dean leaves Google; Hassabis steps back from DeepMind CEO Dean, Google's 30th employee and chief scientist, co-founds public-benefit corporation Discovery Loop with Sanjay Ghemawat, Oriol Vinyals and Quoc Le; Alphabet is a founding investor. DeepMind CTO Koray Kavukcuoglu takes day-to-day leadership while Hassabis becomes DeepMind chairman and Alphabet chief scientist. 4

Nvidia eyes $3B SB Energy stake tied to OpenAI data centers The Information reports Nvidia discussed investing half of the $3B when the Ohio project deal is signed and half at SB Energy's planned IPO, part of roughly $100B in credit support for the campus. SB Energy, also backed by OpenAI, could raise at least $5B in an IPO as soon as next month. 5

OpenAI previews GPT-5.6 Sol at 750 tokens per second via Cerebras Previewed August 13, the limited Ultrafast API tier runs Sol on Cerebras hardware at up to 750 output tokens per second, about 14x OpenAI's Standard tier. Cerebras says Sol finished Humanity's Last Exam in 11h 11m versus 78h 27m for Claude Fable 5, though the measurements aren't independently verified. 6

Google launches Gemini 3.7 Flash with 50% price cut for coding Released August 13 as a refinement of Gemini 3.6 Flash without new pretraining, it handles text, image, audio and video with a 1M-token context. Introductory pricing is $0.75 per million input and $3.75 per million output tokens, half of 3.6 Flash's initial price, rising in January 2027. 7

OpenAI's Jony Ive-designed hockey puck targets smart home in 2027 Reports describe a screenless, battery-powered smart speaker, roughly hockey-puck sized with a hole through the middle, priced around $300–$400 and launching in 2027. It would mark OpenAI's first serious attempt to put ChatGPT into a dedicated household device. 8

Anthropic details Claude text watermarking using Google's SynthID-Text approach Anthropic says it will embed "undetectable" statistical patterns in Claude's text using the SynthID-Text method to comply with the EU AI Act, and it plans a detection API. Light editing probably won't remove the watermark, but a complete rewrite will. 9

Alibaba's Qwen tops 3 billion downloads, half of open-source models Qwen now accounts for more than 50% of global open-source model downloads, with over 200,000 derivative models on Hugging Face. Qwen3.8-Max, a 2.4-trillion-parameter coding and agentic model, is slated for open-weight release this month. 10

Zhipu AI releases GLM-5.3, claims top open-weights coding model The 743B-parameter model's gains come purely from post-training, with coding capability reportedly up 50%; Zhipu says testing uncovered 2,436 vulnerabilities across 269 projects. Weights will be released in about two weeks after security reviews. 11

OpenAI expands ChatGPT ads, updates privacy policy for ad personalization Sponsored content now reaches the US, UK, Mexico, Brazil, Japan and South Korea, with ads served by default in the US (opt-out) and consent-based in the EU/UK. OpenAI says chat context, ad interactions and stored memories can inform targeting, but advertisers don't see conversations. 12