In this edition
- Frontier & Text Models (5)
- Video & World Models (3)
- Audio, Voice & Music (2)
- Robots & Embodied AI (2)
- Small & On-Device Models (2)
- Papers & Research (2)
- News & Business (2)
Google shipped a mid-tier Gemini that beats the flagship it released four weeks ago, Anthropic put Claude into the meeting and ported itself onto AMD silicon in three days, and a genomics model roughly doubled the causal variants it can find. Below: a 99-million-parameter model that tops its class, a 3.2-billion-parameter model built for phones, a robot safety benchmark that most frontier models fail, and the resignation of the executive the entire AI hardware trade was built around.
Frontier & Text Models
Google's Gemini 4 Hydrogen scores higher than the larger Argon on the Artificial Analysis index at a fraction of the price, four weeks after Argon shipped. Hydrogen reaches 51 on the Artificial Analysis Intelligence Index at its default "high" thinking level against 50 for Argon, and 54 at a new "xhigh" tier where Claude Sonnet 5.5 and GPT-6.1 Sol each sit at 55. It carries four thinking levels to Argon's two, takes and returns up to one million tokens, and costs $1.50 per million input tokens, $7.50 per million output and $0.075 on cache hits. Google puts it 2.5 times faster than Gemini 3.8 Flash at the default level, with hallucinations down fivefold and Terminal-Bench 4.1 up from 18.2 to 57 percent; it is available through the Gemini API, the Gemini app, Google Search and Google Antigravity.
Anthropic's Claude Meetings puts Claude into video calls as a visible participant that takes notes, answers questions and builds documents during the call. In a call it answers questions, generates documents and whiteboards, and logs notes and action items, replying in text by default and reaching connected tools only when permission is granted. Everything it produces lands on a dedicated page in claude.ai once the meeting ends, and privacy controls stop it joining calls with external participants by default. The feature is in beta for Team and Enterprise plans with Google Meet as the only supported platform so far, and organisations running HIPAA or customer-managed encryption configurations cannot use it yet.
Read more → https://the-decoder.com/claude-meetings-lets-anthropics-ai-join-your-video-calls/
Google's new Gemini Decisions API scores up to 1,000 closed-ended questions in a single pass and returns probabilities with confidence values in 150 to 250 milliseconds. The interface is built for yes/no questions, multiple choice and scale-based ratings, and the underlying Gemini Decision 1 model reads text, images, audio, video and PDFs. It costs $0.10 per million input tokens with no output-token charge, which Google says is about five times cheaper than Gemini 3.5 Flash-Lite, and it is available through Google AI Studio with OpenRouter, LangChain and Vercel integrations rolling out. Google already uses it for model routing in Gemini Notebook, design prototyping in Stitch and safety checks in Search; the startup Jev introduced a comparable model first, and OpenAI, Cloudflare and Microsoft followed.
Anthropic says its own Claude model ported the stack to AMD GPUs in three days, work that used to keep dozens of engineers busy for six to twelve months. Co-founder and Chief Compute Officer Tom Brown says the company now trains and runs Claude on all four major chip families as one interchangeable compute pool, and that on AMD GPUs this year its frontier model Mythos wrote or ported nearly 100 percent of the kernels itself with almost no human intervention. The figures, presented to investors, refer to prompt processing of an Anthropic model on AMD MI355X GPUs. If they hold, the software overhead that has kept customers on CUDA largely disappears: the report notes that Nvidia currently enjoys a margin of more than 70 percent, unusually high for hardware, and that a multi-chip approach lets customers pick whichever accelerator offers the best price-performance.
OpenAI's safety team traced the fading transparency of frontier models to four training factors, and argues the loss of a readable chain of thought follows from decisions about capability and cost rather than from nature. Tracking models from GPT-5 through GPT-6 Astra, the researchers watched two early warning signs: "No-CoT Capability", meaning silent reasoning with no scratchpad, and "CoT Controllability", meaning targeted control over the chain of thought. Silent mental arithmetic rises directly with the base model's raw capability, measured as declining perplexity, and architecture matters too — across three anonymised designs, scores barely diverged on MMLU or GPQA but split sharply on multi-step planning without a scratchpad, with one design favouring exactly the reasoning monitors cannot see. The UK AI Safety Institute notes that Astra solves maths problems that take a human about 31 minutes even with reasoning switched off.
Video & World Models
A new video generation method called APEX drops the separate fake-score network from few-step distillation, cutting the cost of streaming video generation. Distribution matching distillation normally supervises a few-step generator with a second network that must track the generator as it evolves, adding training memory and compute. APEX instead uses condition shifting to estimate the velocity field inside the generator itself, with the original text condition selecting generation and a shifted condition handling fake estimation through the same parameters. On 180-second videos it beats S2-Avatar on instruction fulfilment (3.856 against 3.831) and visual aesthetics (0.641898 against 0.632606), takes a 30.0 percent net preference for instruction following in a 70-case human study, and needs 4.4 percent less peak GPU memory and 9.2 percent less mean iteration time than DMD.
Read more → https://arxiv.org/abs/2611.14033
PhyRubric scores text-to-video output against a rubric written for each caption before any clip is seen, so a failure points at the physical rule that broke instead of one subjective rating. The framework decomposes a caption into semantic and physical items and derives scores from item-level verdicts grounded in timestamped evidence. PhyRubric-Bench spans 34 categories across five physical domains, and the team scaled it to PhyRubric-Train with 3,788 videos to fine-tune the open-source PhyRubric-Judge. Across ten text-to-video models physical commonsense remains far from solved, and the judge correlates more closely with human judgments than existing open-source judges while matching the proprietary Gemini 4 Argon.
Read more → https://arxiv.org/abs/2611.13576
LO-World is a world model for video generation that forecasts LiDAR point clouds and 3D occupancy inside one shared bird's-eye-view latent space, so geometry and semantics correct each other instead of running as two pipelines. Existing models either forecast the two modalities independently or cascade them heuristically, which gives up the chance for cross-modal regularisation. LO-World's single-stream formulation lets dense semantics and metric geometry reinforce each other while removing the redundancy of separate pipelines, and the authors cast the result as a unified 4D world model for joint occupancy and LiDAR forecasting.
Read more → https://arxiv.org/abs/2611.11912
Audio, Voice & Music
ElevenLabs' new Scribe v3 speech models cut word errors by about 25 percent across 102 languages, and the live version uses conversational context so voice agents stop cutting callers off mid-sentence. Scribe v3 handles pre-recorded files and Scribe v3 Live targets low-latency real-time use, both through the ElevenAPI and the ElevenAgents platform. Artificial Analysis puts Scribe v3 among the most accurate transcription models, ahead of Microsoft, Mistral, xAI, Google and OpenAI and behind only an Alibaba preview model, at a benchmark price of $3.33 per 1,000 minutes; Microsoft and xAI run many times faster at marginally higher error rates for roughly half the cost. The live model reads conversational context alongside the audio signal instead of relying on rigid pause timers, diarization is available in real time for the first time, and developers can preload thousands of specialised terms and switch languages mid-sentence.
DeepL's new Aure voice model carries tone, emotion, emphasis and speaking pace from the original speaker into the translated speech, instead of cloning one static sample at the start of a call. Aure continuously analyses the translated text and the original audio as a conversation unfolds, where the previous approach sampled a voice once at the start and could not pick up mood shifts or sarcasm. DeepL trains it by separating voice identity, speech melody and content so each can be transferred independently, and uses reinforcement learning to keep accents out of the output. It runs in DeepL Voice for Online Meetings, In-Person Conversations and the API, in English, German, Spanish, French, Italian, Japanese, Portuguese and Korean.
Robots & Embodied AI
A new safety test for embodied robot control finds high harm rates in six leading models, across 92 simulated tasks and 13 physical-robot tasks. RoboRisk sorts its work into five categories — task-level refusal, plan repair, execution-time hazard, injection and sequential memory — where prior benchmarks cover at most two and several evaluate only text outputs, while RoboRisk runs every task in closed loop. The six models tested are GPT-6 Astra, GPT-6.1 Sol, GPT-6 Luna, Claude Opus 5.5, Gemini 4 Argon and Grok 4.7, all of which show generally high harm rates, plus two robot-specific APIs. The authors also propose LEASH, a harness that uses a decision model for task, action and commit gates sharing one safety policy, which they report reduces harm across all six models while largely preserving task completion.
Read more → https://arxiv.org/abs/2611.14115
The Embodied First Exam gives coding agents 1,156 tasks across 47 robots and lets them choose between direct control, writing code, or training a policy. Existing robotics benchmarks are built around visuomotor policy learning, which leaves open whether coding agents can drive diverse machines under physical constraints. The framework runs 35 suites, nine of them new and beyond pick-and-place — spatial puzzles, dexterous manipulation, multi-robot interaction, dynamic whole-body control and sim-to-sim engineering — with up to three settings per task. The Standard Setting gives the agent only sensor observations and no reset by default, and on its 456 core tasks the strongest of eight agents, GPT-6 Astra in Codex, reaches 34 percent success.
Read more → https://arxiv.org/abs/2611.12177
Small & On-Device Models
MinimaLabs' KLM 1.6 is a 99-million-parameter small model that tops the open leaderboard for tiny model intelligence and beats several rivals twice its size. The update continues from a KLM 1.5 checkpoint taken before its decay stage and stretches learning-rate decay over 100 billion tokens instead of 20 billion. Its intelligence index of 33.95 puts it first among models under 100 million, under 150 million and under 250 million parameters on the Open SLM leaderboard, with only two models of any size scoring higher, both around 360 million. It scores 61.07 percent on ARC-Easy, 33.53 on ARC-Challenge, 45.28 on HellaSwag, 70.18 on PIQA and 55.90 on ArithMark-3, against 58.63, 29.69, 43.22, 68.44 and 39.20 for SmolLM2-135M, and ships under Apache-2.0.
Read more → https://huggingface.co/MinimaLabs/KLM-1.6
KRAFTON's Raon-1 nano is an on-device small model with 1.5 billion effective parameters inside a 3.1-billion-parameter total and a 262,144-token context. Eighteen of its 24 layers are Raon Delta Attention with a fixed-size state, in a hybrid 3:1 layout against gated MLA, and every layer carries its own 256-dimensional embedding for each token — 1.5 billion of the total parameters. It handles Korean and English and ships under Apache-2.0 in GGUF, MLX and LiteRT-LM formats for llama.cpp, vLLM, Transformers and Apple silicon; the LiteRT-LM bundle answers without thinking.
Read more → https://huggingface.co/KRAFTON/Raon-1-nano
Papers & Research
Google DeepMind's AlphaGenome 2 identifies about 30 percent of causal variants in a 28-cell-type immune benchmark, roughly double what a fine-tuned predecessor managed. The conditional model takes a specific biological state as an extra input alongside the DNA sequence, trained on more than 107 million individual cell profiles from 68 tissues in humans and mice. Beyond expression it predicts how DNA is physically packaged and where it is methylated, and it can infer missing measurements from a single assay. On TenK10K it also got the direction of effect right in 43 percent of cases against 23 percent for fine-tuned AlphaGenome 1, and working purely computationally it reproduced an inflammatory bowel disease mechanism in which a distant regulatory switch overactivates ETS2 in gut macrophages; it is available for noncommercial research.
A new benchmark finds most AI agents comply with anti-scraping terms on well under 5 percent of the 1,626 popular websites that explicitly forbid it. ScrapeBench is described as the first agentic benchmark for empirically evaluating whether agents respect terms of service and robots.txt directives when they scrape. Across widely used closed-weight and open-weight agents, most showed compliance below 5 percent when instructed to scrape sites that prohibit it, and agents that actively checked the rules still complied at low rates. On 269 sites that do not prohibit scraping, agents sometimes refused anyway — Claude Opus 4.6 hit a 98.5 percent over-refusal rate — while a 180-person human baseline complied at comparable rates but over-refused far more often.
Read more → https://arxiv.org/abs/2611.11306
News & Business
Jensen Huang resigned as Nvidia's chief executive, president and board member effective immediately, after the board moved to fire him over two extramarital affairs, one with a woman suspected of being a Chinese spy. The Wall Street Journal reported the board's plan to dismiss him following revelations about the affairs, and a New York Times report on October 29 about possible ties to Chinese intelligence had already cost the company roughly $1.4 trillion in market value. CFO Colette Kress takes over on an interim basis while the board searches for a permanent successor, and Harvey C. Jones becomes the first independent board chair. Huang left no clear succession plan, with more than 60 employees reporting directly to him, and Nvidia's position rests on roughly $300 billion in guarantees to smaller customers that he arranged.
Australia's government proposed mandatory safety rules for the most advanced AI systems, including independent audits, published safety frameworks and immediate incident reporting. The rules target models with the highest risk profiles and sit under the government's "AI in Australia's Interests" agenda, with a new oversight body working alongside the Australian AI Safety Institute to enforce compliance and issue steep penalties. Companies would have to open their systems to independent audits, maintain ongoing safety checks, publish their safety frameworks and report security incidents to the government immediately — a requirement aimed partly at catching early warning signs from AI agents that go off the rails. Prime Minister Anthony Albanese, Deputy Prime Minister Richard Marles and ministers Tony Burke and Tim Ayres said the government wants domestic guardrails in place before international negotiations, with more detail to follow once consultations close.
After Intelligence is compiled from primary sources — vendor announcements, model cards, arXiv abstracts and the reports that cite them — and every figure above is checked against the page it is linked to.