Frontier Model Releases
Gemini 4 Argon. Google globally rolled out Gemini 4 Argon, its new flagship model, available via API/AI Studio and to Ultra/Pro subscribers in Gemini, Antigravity, and Search AI Mode. It launches at $2/million input and $10/million output tokens, doubling December 1, and Google reports top scores on the Vals Index and ERQA robotics benchmark, with early cyber defender access through Fairwind.
- Gemini 4 Argon: Built to solve your hardest problems(opens in a new tab)
- Gemini 4 Argon for Robotics(opens in a new tab)
- Gemini 4 Argon: Built to solve your hardest problems(opens in a new tab)
- Gemini 4 Argon is now available worldwide as Google's most capable frontier model(opens in a new tab)
- Google finally releases Gemini 4 Argon(opens in a new tab)
- llm-gemini 0.35(opens in a new tab)
- Gemini 4 Argon begins rolling out in Gemini app and Antigravity, but only if you pay(opens in a new tab)
- Developing Gemini 4 Argon responsibly and safely(opens in a new tab)
Meta Muse Wonder. Meta released Muse Wonder and Muse Wonder Full, 1M-token multimodal models built for long-running agent work, live on the Meta Model API and Muse Code. The models lead several benchmarks but trail Claude Opus 5.5 on ProgramBench.
Liquid AI LFM2.5. Liquid AI released LFM2.5-24B-A2B, a 24B-parameter reasoning model with only 2B active parameters, 131K context, and explicit chain of thought. It scores competitively against larger agentic models and runs efficiently on-device, reaching 184 tok/s on an Apple M5 Max.
Agent Frameworks & Local AI
Open SWE framework. LangChain released Open SWE, an open-source framework for internal coding agents built on Deep Agents and LangGraph, capturing patterns from Stripe Minions, Ramp Inspect, and Coinbase Cloudbot.
Deep Agents updates. LangChain's Deep Agents added harness engineering improvements that lifted it to Top 5 on Terminal Bench 2.0 via self-verification and tracing, plus context management for long tasks and a beta Deploy service as an open alternative to Claude Managed Agents.
Open models for agents. LangChain reports open-weight models GLM-5 and MiniMax M2.7 now match closed frontier models on file operations, tool use, and instruction following at a fraction of cost and latency.
Everything as MCP. A new workflow reverse-engineers existing apps and has an LLM generate Model Context Protocol wrapper servers, exposing features for agent use and giving total control.
Tencent agent memory. Tencent open-sourced an agent memory system that runs entirely locally on SQLite.
Local hardware notes. Community reports show a local Qwen3.8FN agent on a used Tesla P40 can be cheaper than cloud models at off-peak electricity rates, while running Qwen 27B on a jailbroken PS5 hits ~20 tok/s; 3090 FE owners should cap power and raise fan speed to avoid VRAM overheating above 96C.
Industry & Business
Character.AI bankruptcy. Character.AI filed for bankruptcy and is seeking a buyer; Disney offered $20 million for a non-exclusive perpetual license to some of its technology. If no buyer emerges by early December, the chatbot service could shut down.
- Character.AI files for bankruptcy, to sell some assets to Disney for $20 million(opens in a new tab)
- Disney offers $20M for Character.AI’s scandal-plagued chatbot tech(opens in a new tab)
- Character.ai will shut down if no buyer emerges, capping a fall from $2.7 billion Google deal to bankruptcy(opens in a new tab)
OpenAI security and legal. An OpenAI engineer was arrested for allegedly planting a recording device in an executive conference room, and a judge ordered OpenAI to expedite hiring records to Apple in a trade secrets lawsuit related to former Apple employees in hardware and robotics.
Anthropic custom source control. Anthropic built Scion and Antique, custom source control and code review systems, after daily merges jumped from ~2,500 to over 30,000, with a main branch history near 100GB and 25 million checkouts per day.
Cerebras home AI. Cerebras CEO Andrew Feldman predicted personal AI assistants will soon run locally in home appliances, handling some work on-device and some in cloud, with security perimeters and the ability to make purchases.
Shopify agent commerce. Shopify will let shoppers connect any AI agent to browse and purchase from its catalog of over a billion listings, expanding beyond the Meta Muse partnership with Shop Pay checkout.
Enterprise AI security. Pluto Security raised $42 million to help companies secure AI and vibe-coded apps; the startup says ~60% of risks it detects come from non-developer-built apps.
AI in game development. Microsoft formed Xbox Labs to embrace AI-powered game decompilation and modding, while Epic told developers on an upcoming Disney game to use AI tools like Claude and full AI art workflows.
Orange Business AI platform. Orange Business built Live Intelligence Studio on LangChain and LangSmith, deployed across 120,000 employees and adopted by over 150 European customers.
Applied AI & Robotics
Dyna-2 physical agents. Dyna Robotics deployed Dyna-2.1, a physical robot agent trained on one million hours of human video, with hour-long autonomy and a 99.4% success rate folding napkins for 24 hours straight, including recovery from failures.
Industrial Next retooling. Industrial Next is building AI-driven robotic stations powered by foundation models that let factory production lines retool in minutes instead of weeks, and announced a partnership with BMW Group Technology Office.
On-device scam detection. Google DeepMind and Singapore's GovTech fine-tuned a lightweight Gemma model to run on-device for scam detection, protecting privacy while addressing localized deception techniques.
Wildlife deterrence AI. Flox Intelligence is building AI to keep wildlife away from train tracks, airports, and roads, addressing wildlife strikes that cost aviation millions annually.
Physical AI red teaming. Laplace Intelligence is automating red-teaming for robotics and self-driving cars, stressing physical AI systems to discover failures before real-world deployment.
Vision-restoring BCI. Science Corporation implanted its PRIMA vision-restoring chip in a first Stargardt disease patient, part of its broader push into brain-computer interfaces beyond motor prostheses.
Safety, Policy & Discourse
Scientists demand airgaps. An open letter from 406 scientists including Hinton, Bengio, and Russell calls for AI models to be fully disconnected from the internet until safety improves, citing known containment failures.
Doom rhetoric pushback. UC Berkeley professor Michael I. Jordan dismissed AI existential risk as science fiction and a 'grand moralizing performance,' while Cerebras CEO Andrew Feldman also urged AI leaders to highlight near-term medical benefits instead of doomer narratives.
AI deepfake spread. A viral AI-generated image of actress Mia Goth spread widely despite being fake, prompting her representatives to call Instagram's handling 'incomprehensible' and raising concerns about platform verification.
Consumer & Platform AI
Facebook algorithm control. Meta brought Instagram's Your Algorithm to Facebook, letting users describe feed preferences in natural language and update recommendations in real time, alongside more LLM-based ranking.
Vibe-coded apps on Play. Google now lets users launch private, vibe-coded Android apps through Google Play with a standard Google account, keeping them separate from the public catalog; Ask Play and Game Finder are also expanding AI-powered discovery.
That's everything for this day – about a 5-minute read.