Model Releases
Gemini 4 Carbon. Google made Gemini 4 Carbon generally available after a two-week security-defender test. It is aimed at long-horizon agentic workflows and complex coding, and Antigravity is previewing a '/swarm' feature that turns one prompt into a team of agents.
Anthropic Fable 5.2. Anthropic released Claude Fable 5.2 and Mythos 5.2, the same model with different safeguard classifiers. Fable 5.2 completed 41.3% of real freelance jobs on the Remote Labor Index, up from 2.5% a year ago.
NaiveAI N0.5-Flash. NaiveAI released an open-weights coding model built on Xiaomi MiMo-V2.5, tuned for token speed and long-context reasoning. It claims up to 2,100 tokens per second using its NaiveRT inference engine.
Open Models & Tools
MiniMax H3-Image. MiniMax released H3-Image, an open-weight image generation model built on the H3 architecture, with a Hugging Face repository.
Zyphra ZONOS3-Mini. Zyphra released ZONOS3-Mini, an open-source sparse MoE text-to-speech model with 800M total and 100M active parameters. It supports zero-shot voice cloning and runs on browsers, laptops, phones, and server GPUs.
Bielik-11B v3.1. SpeakLeash released Bielik-11B-v3.1-Instruct, an 11B multilingual model covering 56 languages with an emphasis on Polish and European languages, improved instruction following, tool use, and reasoning.
XeMaya for Intel Arc. XeMaya, a fork of XeStrata, brings DeepSeek V4.1 and GLM 5.3 Flash to Intel Arc GPUs via SYCL and Level Zero, with benchmark results on B-series cards.
Streaming Markdown renderer. A zero-dependency streaming Markdown renderer for AI chat UIs reparses incrementally and repairs unfinished syntax, benchmarking 0.25s on 100k characters versus 2.4s+ for markdown-to-jsx. It adds agent-specific blocks like tabs, stats, and file trees.
Agent Products & Enterprise
ChatGPT Business Agents. OpenAI launched a beta Business Hub where companies can create profiles and verified Business Agents that answer customer questions in ChatGPT. It includes product feeds, comparison shopping, and business insights.
Anthropic office suite. Anthropic added Claude Sheets, Whiteboards, Data, and Interactive Connectors, with Claude Meetings coming soon to join calls and take notes. Enterprise customers get audit logs, isolated VMs, and a compliance API.
- Anthropic turns Claude into a full workspace with Plugins, spreadsheets, whiteboards, and soon Meeting notes(opens in a new tab)
- Anthropic launches Claude Meetings to turn workplace discussions into plans and tasks(opens in a new tab)
- Anthropic has its own AI-powered office suite now(opens in a new tab)
OpenAI Odessia acquisition. OpenAI acquired AI travel agent startup Odessia and hired its founder Francis Davidson; the tool has been shut down and no travel-specific product was announced.
Workday agent platform. Workday built an internal enterprise agent platform for HR, finance, and IT using LangGraph for orchestration and LangSmith for observability. A team of four delivered it under strict audit and data governance.
eBay blocks agents. eBay appears to be blocking third-party AI agents such as Meta's Muse from browsing listings, causing Muse to fail or redirect to mirror sites, the latest platform-agent friction.
Security & Policy
Grok Build leak. Security firm Claroc found that thousands of Grok Build apps were enumerable and exposed confidential data, including two dozen apps built by SpaceXSI itself, because shared links were not truly private.
Dormant neuron backdoors. Microsoft researchers demonstrated 'Lazarus,' a backdoor that edits fewer than ten dormant neurons to make an AI agent execute malicious commands when a hidden trigger appears, without retraining. They recommend model provenance and cryptographic signing.
New York RAISE Act. New York opened registration for the RAISE Act, requiring large frontier AI developers to publish safety frameworks, report critical incidents within 72 hours, and issue model transparency reports. Violations can cost up to $1 million.
Jensen Huang affair. Nvidia CEO Jensen Huang admitted an affair with a woman US officials suspect of Chinese espionage. Experts say the incident could threaten America's AI chip lead and national security.
Robotics & Embodied AI
Ring home robots. Amazon announced Ring Rover Cam and Ring Drone Cam, indoor security robots that use an AI Pilot system to map floor plans and patrol autonomously, arriving starting late 2026 and early 2027.
Gemini Robotics 2.5. Google DeepMind introduced Gemini Robotics 2.5, a vision-language-action model that sorts metal parts with 96% success in Mercedes-Benz pilots and released the ASIMOV-Jailbreak safety benchmark.
Nvidia GR00T 2. Nvidia's GR00T 2 is a world-action model that predicts both robot movement and a short future video, enabling human video demonstrations to transfer across different robot arms and reducing teleoperation.
Research & Benchmarks
Meta e-means. Meta FAIR released e-means, a clustering algorithm that matches k-means loss in 1.4-2.6 epochs instead of 25 and provides 10-45x GPU speedups. It is available in FAISS 1.16.0.
Deletion Tuning. A paper finds that zeroing fewer than 0.05% of a model's weights can match full fine-tuning across RL, SFT, and distillation regimes, while forgetting less and producing tiny checkpoints.
Agent distraction. Google research shows coding agents get derailed by stale TODO comments, old bug references, and outdated specs, causing wasted tokens or incorrect logic. They propose cross-referencing TODOs with bug status.
PSI quantum result. Using early access to Anthropic's Fable 5.2 model, startup lab PSI solved the 18-year-old Aaronson-Ambainis conjecture, with the proof autoformalized in Lean.
Qwen3.8 less-common languages. A benchmark of 202 GitHub tasks in 13 less-common languages found Qwen3.8 27B solves 43% of tasks versus Qwen3.6's 17%, though the newer model helped select hard tasks, biasing results down.
That's everything for this day – about a 5-minute read.