- Models
Claude now hides a watermark in the text it writes. According to an O'Reilly Radar analysis, Anthropic's support page says its newest models and some older ones watermark text on every Claude surface worldwide, a step tied to the EU AI Act's machine-readable marking requirement.15 Neither Anthropic, OpenAI nor Google lets the public check text against its detector.15
- Models
Baseten released bison-1 and groundhog-1, open decision models built on Gemma 4. Baseten says bison-1 ranks first of 19 models on the Artificial Analysis Decision Index v2.1 with a score of 76. Both models take text and images with a 128k context and cost $0.10 and $0.035 per million tokens.16, 17 Requests longer than the context limit are refused rather than truncated, so pipelines need a length check.16
- Industry
Nvidia's interim CEO Colette Kress put new limits on financing for AI cloud partners. On her first earnings call as interim CEO, Kress reported a $115.8 billion quarter. She said future help for partners will focus on land, power and shell, and that commitments to neoclouds apply only if operators secure their own financing and deploy the capacity.18 Nvidia's maximum exposure across infrastructure guarantees reached $118 billion at the end of the October quarter.18
- Policy & legal
Cursor is weighing a move away from Chinese open-weight models, a House report says. A joint report from two House committees says the SpaceX-owned coding tool is considering phasing out Chinese models, and notes that its Composer 2 was built on Moonshot AI's Kimi K2.5.19 The report recommends banning models developed in China from federal systems and federal contract work.19
- Models
RadixArk released DaisyRose-27B, an agentic coding model trained with reinforcement learning only. The Apache 2.0 model is post-trained from Qwen3.8-27B and more than doubles its Terminal Bench 4.1 score, from 5.6 to 12.0. Training ran 18 days on 144 H200 GPUs.20 It fits on one GPU, at about 52 GB in BF16 and 23 GB in NVFP4, and all reported scores are self-reported.20
- Industry
Signal now blocks screenshots by default to keep OS-level AI agents out of chats. Signal's new Block Siri AI setting on iOS 27 is on by default, and on Android the app turns on Screen security when it detects an assistant reading the screen.21, 22 President Meredith Whittaker called the fix incomplete, said Signal cannot do the same on macOS, and asked Apple, Google and Microsoft to give developers a documented way to opt out of AI data access.21, 23
- Models
Nvidia released Nemotron-Labs-Sigma, a family of continuous diffusion language models. The 14 open checkpoints at 3B and 8B scale denoise text in a 16-dimensional embedding space. Nvidia reports that the 8-step distilled 8B model keeps a 70.34 average on its six-task suite.24 The self-speculative variants ran 3.46 times faster than autoregressive decoding on a GB200 with SGLang, by Nvidia's own measurements.24
- Policy & legal
House data shows US officials approved nearly 900,000 AI chips for Chinese companies. A Commerce Department summary dated May 11 and released by the House Foreign Affairs Committee shows 879,868 Nvidia and AMD chips approved out of applications for 3.44 million. Approved buyers included Tencent, Alibaba and Baidu.25 The Washington Post reported that almost 580,000 of the approved chips were Nvidia's, and Nvidia said its H200 shipments were legal and authorized.26, 25
- Products
OpenAI brought ChatGPT Space to Plus subscribers. Space is a workspace for documents, slides, spreadsheets and sites, and it can automatically refresh pages using connected apps such as Gmail and Slack on a schedule.27 It was previously limited to Pro, Business and Enterprise plans.27
- Products
WiCi released a free desktop app that runs 800B-parameter models by streaming weights from the SSD. WiCi Infer is built on the open-source InferSSD engine, a fork of llama.cpp. It keeps routed expert weights on disk, needs about 520GB of free SSD space and at least 16GB to 24GB of memory.28 WiCi's own benchmarks show about 2.7 tokens per second for DeepSeek V4.1 Flash on a 48GB M5 Max MacBook Pro.28
- Funding & deals
After its IPO, Anthropic faces the question of what follows Claude Code. Business Insider reports that Anthropic, which closed its first trading day valued at $1.7 trillion, now needs a business beyond AI coding, and points to biotech, cybersecurity and robotics as possible areas.29
- Research
METR updated its time horizon chart and found models still fall short on judgment-heavy tasks. METR says time horizons on its combined task suites are doubling roughly every 3.6 months.30 When asked to design AI evaluations, though, no model produced one that METR staff found useful for their research.31 Newer models such as Opus 5.5 reached what METR calls borderline above the bar on that task.31
- Products
OpenAI is testing a redesigned ChatGPT mobile app with a dedicated tab for dots. The new layout adds a Home tab, a Dot tab and a grid of tools such as Codex, Space and Projects that users can pin to the tab bar. It is rolling out first to eligible Pro users, outside the EEA, the UK, Switzerland and South Korea.32, 33
- Industry
A hosting breach aimed at AI provider keys is pushing firms to treat agent keys like user logins. An Accounting Today column describes a cloud hosting platform disclosing that attackers sought AI provider API keys, and urges firms to scope, rotate and log each agent's keys under the FTC Safeguards Rule.34
- Products
Google's Gemini can now generate GIFs. The looping animations run about three seconds and use the Gemini Omni 1.2 Flash video model, initially in seven countries including the US, UK and Japan.35
- Models
OpenAI introduced MedChemBench to test AI reasoning in medicinal chemistry. The benchmark has 99 problems drawn from 117 papers and patents, and GPT-6.2 Astra scored 84.5%.36, 37 OpenAI says the results do not show that models can discover drugs on their own.36
- Infrastructure
A lawyer's $9,450 server built from 12 used V100s now runs every DeepSeek V4.1 Flash expert at NVFP4. With a third board of cards, all 15,360 experts fit on the GPUs, and the time to the first code token at 256K context fell from 108.4 seconds to 58 seconds.38