Morning Edition · Thursday, November 12, 2026
Tech
Anthropic's Claude ban now reaches staff of China- and Russia-controlled companies wherever they work
The revised Supported Regions Policy takes effect November 12. It covers Claude.ai, Claude Code, Cowork, the API and Claude models resold through cloud platforms, so checks on ownership and location now determine who can buy access to the model.
Markets
Anthropic markets 825 million shares at $70 to $78 for an initial public offering valued at up to $1.4 trillion
Pricing is expected November 18, one day after Nvidia reports earnings. Google, Nvidia and five other strategic partners have agreed to buy $26 billion of non-voting stock at the offering price.
Tech
Claude Fable 5.2 raises Anthropic's FrontierCode score to 49.1% and cuts API prices by 20%
Mythos 5.2 uses the same underlying model with fewer cybersecurity and biology safeguards. It scores 89.6% on CVE rediscovery and is available only to verified organizations.
Tech
AlphaGenome 2 finds 30% of causal variants on the TenK10K test, nearly double its predecessor
DeepMind's new model predicts a variant's effect one cell type at a time from a 1-megabase (about 1 million DNA base pair) window of sequence. It was trained on more than 107 million single-cell profiles and is now in private preview for commercial research.
Tech
Gemini Robotics 2.5 sorts metal parts with 96% success on Mercedes-Benz assembly lines
DeepMind reports success rates 30% to 100% higher than the prior version across more than 200 tasks. The model goes only to partners including Apptronik, Agile Robots and Boston Dynamics.
Tech
Nvidia's 2-billion-parameter GR00T 2 learns chemistry-lab tasks from a handful of demonstrations and will ship open weights
Jim Fan, Nvidia's director of robotics, says a lab demo that ends with the robot cleaning up was run autonomously, uncut and at real speed on onboard compute. Weights are due in December.
Tech
Google prices Gemini Decision 1 at $0.10 per million input tokens and charges nothing for output
The new Decisions API returns calibrated probabilities over yes-no, multiple-choice and rubric answers in a single pass, with 150 to 250 milliseconds of end-to-end latency according to Google.
Tech
Five AI models chatted with 1,277 dating-app matches for six days, and only eight people asked whether they were talking to a bot
The DateBench project says 355 matches sent their phone numbers without having agreed to take part. Separately, Australia has moved to create an AI safety regulator after an OpenAI agent accessed a Medicare portal.
Tech
Claude helped Anthropic's biology lab produce the first 3D structure of a vault particle from a prokaryote
Seven of 15 candidate proteins assembled into vault-like particles, and cryo-electron microscopy confirmed four. Anthropic also lists a Claude-written protocol error that cost about two weeks.
Tech
A guide to running Gemma 4 locally with its refusals stripped out has been saved 14,000 times
Hugging Face model cards for "abliterated" Gemma 4 builds report refusal rates on harmful prompts falling from 99 in 100 to 26 in 100, achieved by editing the weights with no retraining.
Tech
Zyphra releases ZONOS3-Mini, an Apache-licensed voice-cloning model with 100 million active parameters that runs in a browser
The sparse mixture-of-experts text-to-speech model clones a voice from a sample of up to one minute and runs in real time on phones and laptops, and Zyphra ranks it first among open models on its own benchmarks.
Tech
Boston Dynamics fits Atlas with scooter wheels to push its leg motors to their limits under reinforcement learning
A whole-body policy trained in MuJoCo, combined with refined motor models, carried the wheeled humanoid through sideways dashes and a 360-degree spin jump. The company says the build will not be sold.